ECONOMIC ANALYSIS GROUP DISCUSSION PAPER … · The Impact of Uncertainty and Sunk Costs on Firm ... the U.S. Manufacturing Sector by Vivek Ghosal* EAG 01-1 February 5, 2001 ... industry

ECONOMIC ANALYSIS GROUP DISCUSSION PAPER

The Impact of Uncertainty and Sunk Costs on FirmDynamics and Industry Structure: Evidence fromthe U.S. Manufacturing Sector

by

Vivek Ghosal*

EAG 01-1 February 5, 2001

EAG Discussion Papers are the primary vehicle used to disseminate research from economists in theEconomic Analysis Group (EAG) of the Antitrust Division. These papers are intended to inform interestedindividuals and institutions of EAG’s research program and to stimulate comment and criticism on economicissues related to antitrust policy and regulation. The analysis and conclusions of EAG Discussion Papersrepresent those of the authors and do not necessarily reflect the views of the Department of Justice.

Information on the EAG research program and discussion paper series may be obtained from Russell Pittman,Director of Economic Research, Economic Analysis Group, Antitrust Division, U.S. Department of Justice,BICN 10-000, Washington, DC 20530, or by e-mail at [email protected]. Comments on specificpapers may be addressed directly to the authors at the same mailing address or at their e-mail address.

Recent EAG Discussion Paper titles are listed at the end of this paper. To obtain a complete list of titles orto request single copies of individual papers, please write to Janet Ficco at the above mailing address or [email protected] or call (202) 307-3779. Beginning with papers issued in 1999, copies of individualpapers are also available from the Social Science Research Network at www.ssrn.com.

* Contact information: Vivek Ghosal, Economic Analysis Group, Antitrust Division, U.S. Department of Justice,600 E. St., NW, Suite 10-000, Washington, D.C. 20530. Email: [email protected]. The views expressedin this paper are not purported to represent those of the U.S. Department of Justice.

The Impact of Uncertainty and Sunk Costs on Firm Dynamics and Industry Structure:Evidence from the U.S. Manufacturing Sector

Vivek Ghosal

Abstract

Motivated by relatively recent theoretical contributions on the role played by uncertainty, sunk costs andtechnological change in influencing entry, exit and industry structure, we conduct an empirical analysiscovering 267 U.S. manufacturing industries over the period 1963-1992. The key findings are: (1) Greateruncertainty has an adverse impact on the number of small firms and establishments in an industry. Thiseffect is most pronounced in the high sunk cost industries. The larger establishments are generallyunscathed by greater uncertainty. Uncertainty and sunk costs increase concentration and affect firm(establishment) size distribution; and (2) Technological change reduces the number of establishmentsand firms in an industry. The reduction occurs primarily in the small size category. Apart from sheddinglight on the underlying theories, our findings on firm turnover and industry dynamics may haveimplications for research in several areas such as models of creative destruction, competition policyanalysis, and the study of specific industries such as the electric industry which is characterized by a highdegree of uncertainty and is undergoing transformation following deregulation. Further, if entry and exitcan be viewed as the bigger picture reflecting economic activity in industries, our findings may provideinsights into the factors affecting fluctuations in employment and investment.JEL: L11, D80, O30, G10, L40.Keywords: Firm dynamics, Industry structure, uncertainty, sunk costs, information asymmetry, financingconstraints, technological change.

Vivek GhosalEconomic Analysis GroupAntitrust DivisionU.S. Department of Justice600 E. St., NW, Suite 10-000Washington, D.C. 20530.Email: [email protected]

Commenting on the “churning” of firms, Sutton (1997.a; p.52-54) writes that fluctuations in industry profits1

influence entry and exit, and (p.53) “A new attack on this problem has been emerging recently, following the workof Avinash Dixit and Robert Pindyck (1994) on investment under uncertainty. Here, the focus is on analysing thedifferent thresholds associated with entry decisions which involve sunk costs and decisions to exit...” He goes onto note that empirical implementation appears to be both difficult and lags theoretical advances.

While a relevant consideration, we do not address the risk-preference channel; see Appelbaum and Katz2

(1986) and Ghosal (1996) for some discussion of this literature.

1

1. Introduction

Myriad forces regulate entry and exit, and shape industry structure. This paper focuses on three channelsthat could play an important role in influencing the dynamics of industry structure. First, the “optionvalue” literature (e.g., Dixit, 1989; Dixit and Pindyck, 1994) shows that, under uncertainty and sunkcosts, the entry (exit) trigger price will be greater (lower) than the conventional Marshallian entry (exit)price. This results in a wider zone of inaction where entrants (incumbents) are less likely to enter (exit).Second, the literature on “asymmetric information” (e.g., Greenwald and Stiglitz, 1990; Gale andHellwig, 1985) shows that greater uncertainty is likely to exacerbate imperfect information problemsin capital markets and constrain access to credit for some firms. This may affect decisions of entrantsand incumbents who are most likely to be adversely affected by information asymmetries. The abovetwo channels form the main focus of our study. Third, the literature on “technological change” (e.g.,Jovanovic and MacDonald, 1994; Klepper and Simons, 2000) shows that incremental and majortechnological change are likely to affect entry and exit and alter industry structure.

This paper presents an empirical examination of the role played by these channels on thedynamics of industry structure, with our primary focus being on uncertainty and sunk costs. Our studyis motivated by the fact that while the theory linking the option value and asymmetric informationchannels to firm dynamics and industry structure is relatively well developed, empirical evaluation ofthese models seems to be limited. We also examine the significance of technological change as recent1

studies suggest it to be an important determinant of industry turnover. While there are several studiesthat typically provide insights using data on selected industries, we conduct an econometric evaluationthat provides broad insights covering a large number of industries over a long time period.

Our study could be useful for several reasons. For example, our empirical findings may shedlight on the alternate forces that influence the evolution of industry structure and the well documentedskewed size distribution of firms (e.g., Simon and Bonini, 1958), and have implications for theorizing.Second, since the dynamics of entry and exit shapes industry structure and regulates the competitivenessof markets, our findings may provide guidance for public policy issues such as antitrust. Third, the highrates of job creation and destruction (e.g., Davis, Haltiwanger and Schuh, 1996) and the significantintertemporal volatility of investment spending are well documented. If entry and exit can be viewed asthe bigger picture reflecting economic activity in industries, our empirical findings may provide insightsinto some of the driving forces affecting investment and employment.

The paper is organized as follows. Section 2 focuses on the option value and asymmetricinformation channels, outlines the impact of uncertainty and sunk costs on industry dynamics anddiscusses the empirical framework. The role of technological change is examined in Section 3 along2

with links to our empirical analysis; this section also reviews some of the other determinants. Ourempirical analysis uses an extensive database containing information on the number of establishments

Several papers study firm dynamics assuming firm-specific uncertainty; e.g., Hopenhayn (1992), Lambson3

(1991), Pakes and Ericsson (1998). The latter evaluate models of firm dynamics under active v. passive learning;these can be properly subjected to empirical evaluation using only micro-datasets, as in Pakes and Ericsson.

In the simplest settings, the models consider uncertainty about prices holding constant input costs and4

technology. But Caballero and Pindyck (1996) and Dixit and Pindyck (1994), for example, discuss uncertaintyabout cash-flows, profits, among other variables. We return to this in Section 4.3.

2

and firms in an industry, the number of establishments by size class, the degree of industry concentrationand alternate proxies for sunk capital costs. We use industry time-series data to construct measures ofprofit uncertainty and technological change. The data and measurement issues are presented in Section4. Our econometric analysis uses a dynamic panel data model to evaluate the role played by uncertainty,sunk costs and technical progress on the evolution of various industry structure characteristics. Theempirical model and econometric issues are outlined in Section 5, and the estimation results arepresented in Section 6. Concluding remarks appear in Section 7.

2. Uncertainty and Sunk Costs

This section focuses on the central objective of the paper: the role played by the option value andasymmetric information channels. We also outline the links to our empirical analysis.

2.1. Option ValueFirst, we review the theoretical insights from Dixit (1989), Dixit and Pindyck (1994) and Caballero andPindyck (1996). These papers provide a broad framework to structure our empirical analysis usingindustry data. Second, we present links to our empirical analysis. Dixit (1989) models a price-taking3

risk-neutral firm with access to a given technology. Let K be the sunk entry capital investment; r the costof capital; L the exit cost; and Z the stochastic element, which can be couched in terms of any relevantvariable. The conditional standard deviation of Z, FF(Z), measures uncertainty. Production occurs4

incurring variable costs W. Let M and M be the Marshallian entry and exit thresholds: M =(W+rK)H L H

and M =(W-rL), where M is the annualized full cost of incurring the investment and producing output.L H

For M <P<M , a potential entrant (incumbent) does not enter (exit) the industry; this is the MarshallianL H

zone of inaction, relevant in the absence of sunk costs and uncertainty. Let P and P be the entry andH L

exit trigger prices: in the range (0,P ) the potential entrant holds on to its option to enter, and in the rangeH

(P ,4) an incumbent remains in the industry. Uncertainty and sunk costs imply an “option value” ofL

waiting for new information and this raises (lowers) the entry (exit) trigger:

P >(W+rK)/M , and P <(W-rL)/M (1).H H L L

Entry is delayed as entrants require a premium over the Marshallian price. Exit is delayed as theincumbent knows it has to re-incur sunk costs upon re-entry. Thus, uncertainty, F(Z), and sunk costs,K, increase (P /M ) decrease (P /M ) and widen the zone of entry/exit inaction. Dixit’s analytical resultsH H L L

and numerical solutions highlight several interesting results and, for our empirical analysis, we note thefollowing. Given F(Z) (>0), increase in K widens the zone of inaction-Figure 1(a). Similarly, given K(>0), increase in F(Z) widens the zone-Figure 1(b). Numerical results indicate that widening of the gapbetween (P /M ) and (P /M ) is more dramatic for increase in K (given F(Z)) than for increase in F(Z)H H L L

The equations for the optimal solutions are non-linear and do not have analytic solutions. Dixit (1991, p.147-5

48) presents the approximate expression for the range of inaction in the case of small sunk costs:Rn(P /P )=2{[F (K+L)]/W} , where 2 is a constant. From this we can see the impact of changes in K on the zoneH L 2 1/3

of inaction, given L and W. Numerical results in Dixit (1989) show that for a given level of uncertainty F (>0),as K60, dP /dK6(+)4 and dP /dK6(-)4; i.e., even small sunk costs widens the zone of inaction.H L

An adverse demand shock decreases prices along the existing supply curve, whereas a large enough favorable6

shock may induce entry or expansion by incumbents, shifting the supply curve to the right and dampening the priceincrease. Thus, while the upside is truncated by entry, the downside is unaffected. This asymmetry reduces theexpected payoff from investment resulting in the entry trigger price exceeding the Marshallian price. If exit orabandonment is likely, it would create a price floor making each firm more willing to accept a period of losses.

3

(given K). If both F(Z) and K increase, the gap between (P /M ) and (P /M ) widens rapidly.H H L L 5

Caballero and Pindyck (1996) model the intertemporal path of a competitive industry assumingfree entry. Potential entrants’ forecast of future cash flows determines the entry decision. Failure (exit)is exogenous and modeled as a Poisson arrival with the intensity (. Let F be the sunk entry capital cost,A (t) the output of the i firm, A(t) the average industry productivity, a (t) the productivity of the i firmi i

th th

relative to the industry with A(t)=A(t)a (t), and N(t) the number of firms. Industry demand is given byi i

P(t)=M(t)Q(t) , where M is a stochastic industry shock. The firm’s marginal revenue product of capital-1/0

equals P(t)A(t). The distribution of future marginal profitability of capital is asymmetric; during positivei

shocks entry truncates the distribution for the aggregate component of P(t)A (t) - i.e., P(t)A(t) - fromi

above, whereas negative shocks imply larger reduction in marginal profitability. Each firm’s value of6

output is linear in the output-specific stochastic variable, implying that firm-specific shocks areeliminated, except for the exogenous failure rate (. Let B(t)=P(t)A(t) be the average firm’s value ofoutput. Since N(t) is endogenous, it regulates B(t); B(t) is modeled as a regulated geometric Brownianmotion that remains at or below a fixed upper boundary U, the entry point. Let F be the standardb

deviation of the rate of growth of cash flow per firm over the industry. A key result is that M(U/F)/MF >0:b

sunk costs and industry-wide uncertainty cause the entry (investment) threshold to rise above theneoclassical cost of capital. Using data on U.S. SIC 4-digit manufacturing industries, they find the entry(investment) trigger to be much greater than the cost of capital.

Finally, we turn to imperfect competition. First, Dixit and Pindyck (1994, p.309-315) show thatoption value of waiting remains under uncertainty and sunk costs, but an oligopolistic setting requiresconsideration of pre-emption by rival(s) and this may necessitate a faster response. Let there be twopotential entrants F(1) and F(2) and assume, arbitrarily, that F(1) is the leader and F(2) the follower. LetP(1) be the price at which F(1) enters, and P(2) for F(2) enters with P(2)>P(1). If P>P(2) then F(2) entersimmediately, otherwise it waits for price to rise above P(2). P(1) exceeds the Marshallian entry triggerdue to sunk costs and uncertainty. Let F(1) earn profits A(1) during the interval of its sole occupancy inthe industry; when P(1)<P<P(2). F(1) rationally expects profits to fall when F(2) enters. Entry by F(2)will truncate the upper end of the distribution of profits; this is similar to the mechanism described inCaballero and Pindyck. Thus, F(1) must receive a premium before it enters. In this setting there is noexplicit mechanism determining the leader/follower: who enters and when depends on the underlyingconditions. Now consider simultaneous decision making and P=P(1)+, (, is small positive). Price isabove the entry threshold and since neither firm wants to wait, for the fear of being preempted by itsrival and assuming the leadership role, we could get simultaneous entry. Here, entry could be faster thanin the leader-follower setting. Second, there is a large literature (see, e.g., Appelbaum and Lim, 1985;Dixit, 1980; Spencer and Brander, 1992) which, for oligopolistic markets, examines the incentives for

See Audretsch (1995), Dunne et al. (1988), Evans (1987), Klepper (1999) and Sutton (1997.a). In Audretsch7

(p.73-80), the mean size of the entering firm is 7 employees and varies from 4 to 15 across 2-digit industries.Audretsch (p.157-160) shows that 19% of the exiting firms have been in the industry for less than 2 years with amean size of 14 employees; considering exiting firms of all ages, the mean size is 23. Dunne et al. note (p.503):“On average, 38.6% of the firms in operation in each industry in each Census year were not producing in thatindustry in the previous Census...the entrants in each year are responsible for approximately 15.8% of eachmanufacturing industry’s output.” Thus, while the number of entrants is large, their size is very small relative tothe incumbents. Data indicate a similar pattern for exiters. As Martin (1993, p.197) notes, “...new firms are new,small, and (for the most part) doomed, while old firms are larger and are much more likely to survive.”

The data in Dunne et al. (1988) for U.S. 4-digit manufacturing industries over the Census years 1963-19828

(similar to ours) show a raw correlation between entry and exit rates for a given Census period to be between 0.18and 0.33. But, after sweeping out industry fixed-effects, the correlations become negative and range from -0.028to -0.249. They conclude (p.509): “entry and exit patterns are positively correlated across industries, but canbecome negatively correlated when industry-specific effects are removed.” While correlations in the raw data arepositive, the relatively low value indicates considerable cross-industry variation in net entry patterns. Geroski(1995) presents findings for other countries. Finally, there have been numerous studies using net entry data; see,e.g., some of the papers in Geroski and Schwalbach (1991) and the reference there.

4

incumbents or first-movers to engage in strategic pre-commitments to erect entry barriers. Results showthat, under a variety of settings, it is optimal for the incumbent/first-mover to pre-commit. But underuncertainty, the optimal pre-commitment is lower due to greater uncertainty about the success of theentry-deterring strategy. Overall, oligopolistic settings highlight the dependence of outcomes on precisemodel assumptions and it appears difficult to make clear predictions.

2.1.(a). Links to Empirical AnalysisWe start by noting two issues. First, the theory above assumes symmetry in firm size. But industriestypically have a skewed distribution of firm sizes (Simon and Bonini, 1958; Ijiri and Simon, 1977;Sutton, 1997.a). Our manufacturing industry data (Section 4.1) shows this to be the pre-dominantcharacteristic. Further, larger firms are usually older with higher survival rates. The typical profile is onewhere entrants start small and build up to optimal scale. Failure rates are high and exit data shows thetypical exiting firm is small and young. We attempt to incorporate the small versus large firm issue in7

our empirical analysis. Second, our data are on the total number of establishments and firms in anindustry. Thus, we can only study “net entry”. This data-induced limitation implies that we are unableto observe whether uncertainty and sunk costs have potentially differential effects on entry and exit.Several studies have noted a positive correlation between entry and exit rates, implying that net entrydata masks some of the underlying turnover. However, the data also indicates cross-industry variationin patterns of net entry. In Section 4.1 we show that our data contains a reasonable amount of “within-8

industry” variation in net entry, which is encouraging from the perspective of our empirical examination.We summarize the implications for the dynamics of industry structure as follows:

(A) Net entry. Let P and P be the entry and exit trigger prices, FF(Z) uncertainty and SS(K) a measureH L

of sunk entry capital costs. The following basic relationships emerge from theory:(i) MP /MF>0; (ii) M P /MFMS(K)>0; (iii) MP /MF<0; and (iv) M P /MFMS(K)<0.H 2 H L 2 L

From (i) and (iii), greater uncertainty reduces both entry and exit; the effect on “net entry” is ambiguous.From (ii) and (iv), as uncertainty and sunk costs increase, both entry and exit fall; the effect on net entryis ambiguous. Given that, under general conditions, we cannot say whether entry or exit is consistently

Of course, 1 and 2 won’t hold for every industry, but observations from references above appear to make them9

reasonable in general. Considering alternate scenarios, it is possible to argue, for example, that the magnitude ofsunk costs are not too high in large firms. Larger firms are often multi-product and multi-establishment and thesefeatures may make some types of capital investments of large firms relatively less sunk because (i) specificity isless of a problem for a firm that employs capital in many different lines, and (ii) internal redeployment of capitaldoes not suffer from the Akerlof-like “lemons” problem that would arise if the firm tried to sell capital to anotherfirm. Since specificity and lemons problems are among the reasons for the irreversibility of investments it couldbe argued that the degree of sunkenness of capital for large firms is less than it would be otherwise. However,having said this, it still seems difficult to argue in general that sunk costs are lower for the larger older firms.

Using terminology from international trade, larger firms are more likely to show (exit) “hysteresis”; see10

Baldwin and Krugman (1989) and Dixit (1989) on entry and exit of foreign firms following exchange-rate shocks.

5

affected more, the prior prediction would appear to be one of little or no change -inaction- in net entryfollowing greater uncertainty and sunk costs.

Can we predict whether uncertainty and sunk costs affect small versus large firms differentially?We noted that larger incumbents are typically older with higher survival rates. In addition: (a) Largerfirms, being older, are likely to have made greater investments in advertising and distribution networks.Sutton (1991) provides examples in the context of advertizing. These investments, which entail sunkcosts, may erode upon exit and would have to be re-established if the firm re-enters in the future; and(b) R&D involves sunk costs (Sutton, 1997.b) and larger firms are likely to have made greater R&Dinvestments. Exit would entail significant loss of human and physical capital related to product andprocess innovation. If (a) and (b) are reasonable assumptions, then larger incumbents may show more“inaction” regarding exit during periods of increased uncertainty. Using (a) and (b), smaller incumbents9

may be expected to show relatively (compared to larger firms) less inaction regarding exit.10

Regarding entrants, we typically don’t observe significant asymmetry in firm sizes; most aresmall and the successful one’s grow bigger over time. Thus we expect uncertainty and sunk costs toretard entry across the board. Incorporating the firm size distribution issue of incumbents and entrants,and the nature of investments in advertising and R&D, it may be reasonable to conclude the following:(i) For industries with high F(Z) and high S(K), both entry and exit of smaller firms will be relativelylower. With low F(Z) and low S(K), both entry and exit of smaller firms will be relatively higher. Thusno clear prediction emerges regarding net entry; and (ii) Given the large firm issues discussed above,we predict greater inaction in large firm net entry data. Thus, even after addressing the small versus largefirm issue, it appears that we cannot make clear predictions regarding net entry and changes in thenumber of firms and establishments over time.(B) Size distribution. If greater uncertainty and sunk costs lower entry and exit, and we cannot clearlyassess asymmetry in effects on small versus large firms, then the effect on size distribution is unclear.(C) Concentration. Given the above arguments, the effect on industry concentration is ambiguous.

The broad framework for our empirical analysis is as follows: (i) Create alternate measures ofindustry-specific sunk costs (Section 4.2) using the insights in Kessides (1990) and Sutton (1991). Thesemeasures are used to create low versus high sunk cost industry groups; (ii) Construct industry-specificmeasures of uncertainty (Section 4.3); and (iii) Examine the impact of uncertainty and sunk costs on thenumber of (small and large) firms and establishments and industry concentration. While we are unableto examine fluctuations in entry and exit separately, our study will provide an understanding of howuncertainty and sunk costs affect the broad structure of industries as measured by the number of firms

Ghosal (1996) finds that greater uncertainty decreases the number of firms in an industry. But this cross-11

section study does not shed light on the dynamics of industry structure or firm size distribution.

See Fazzari et al. (1988), Greenwald et al. (1984), Myers and Majluf (1984) and Stiglitz and Weiss (1981).12

6

and establishments, and concentration, and may shed light on the underlying theories.11

2.2. Asymmetric InformationWhile numerous papers have focused on the option value channel, another potentially important channelhas received less attention; asymmetric information and financing constraints. We focus on the paperby Greenwald and Stiglitz (1990) as it reviews the literature and presents new results. While the modelis couched in terms of investment, there are implications for entry and exit. Greenwald and Stiglitzconsider a firm where decision makers maximize expected end-of-period equity minus an expected costof bankruptcy; the latter plays a key role in decision making. At first they assume firms use circulatingcapital, and inputs must be paid before output prices and revenues are realized. Production occurs viaCRS technology. Firm equity level in period ‘t’ is assumed to equal the level in period ‘t-1' plus profits.Finally, they consider the scenario where some firms are equity constrained whereas others areborrowing constrained. For example, one can think of an industry with a distribution of firms, with somefirms falling into each category. Incorporating these features, the investment function is:

I =h(w , r , F , a , b ) (2),t t t t t-1 t

where w and r denote wages and expected return to borrowers, F is uncertainty about profits, a is thet-1

equity level from last period, and b is maximum level of borrowing; h , h , h <0 and h , h >0. Investmentw r F a b

depends on wages and interest rates, but also on the equity level, uncertainty and borrowing constraints.Finally, they extend the circulating capital model to incorporate investment in long-lived physicalcapital; the basic insights from (2) carry over. The key results relevant for our analysis are as follows.Greater uncertainty about profitability: (i) increases the absolute and incremental risk of bankruptcy atany level of investment since they cannot absorb the increased risk by issuing more equity-this lowersinvestment; and (ii) exacerbates borrowing constraints (for certain firms) and reduces investment. Thisfollows as some firms are more likely to face borrowing constraints due to information asymmetriesbetween borrowers and lenders. Therefore, the impact of increased uncertainty on investment will12

differ across firms, depending on the extent of financing constraints. Gale and Hellwig (1985), withinthe context of a somewhat different model, derive results similar in spirit to those above.

Next we focus on sunk costs. Lenders are more likely to be reluctant to provide financing ifcapital embeds large sunk costs. If asset specificity is high then capital is likely to have low resale value;collateral has less value. Regarding asset recovery by debt holders, Williamson (1988) writes:

(p.571) “Of the several dimensions with respect to which transactions differ, the most importantis the condition of asset specificity. This has a relation to the notion of sunk cost...” (p.580) “Inthe event of default, the debt-holders will exercise pre-emptive claims against the assets inquestion....The various debt holders will then realize differential recovery in the degree to whichthe assets in question are redeployable...the value of a pre-emptive claim declines as the degreeof asset specificity deepens...”

Gertler and Gilchrist find that tightening of monetary policy affects real activity in small firms more than13

large firms; they use this to argue that firm size is a proxy for the ability to access capital markets. Fazzari, et al.(1988) find small firms are more reliant on internal cash-flows to finance investment. They interpret this asevidence that small firms are more financially constrained. Ramey (1993, p. 7-8) writes that the small v. large firmresults offer “very compelling evidence in favor of the hypothesis that there are credit market imperfections.”

In a different context, Chan and Chen (1991), using market value of equity as a proxy for size, show that the14

same bad news affects the return of a portfolio of small firms significantly more than larger firms. This imposesgreater restrictions on small firms’ access to capital. They attribute this to imperfect information in capital markets.

7

Shleifer and Vishny (1992) develop on this theme to formulate a general equilibrium model of assetsales. In their model asset specificity is an important determinant of leverage and helps explain cross-industry and intertemporal patterns of financing; for example, the ease of debt financing is inverselyrelated to the degree of asset specificity. Thus, greater sunk costs are likely to exacerbate asymmetricinformation problems and tighten borrowing constraints.

While the above arguments are broadly couched in terms of investment decisions, they haveimplications for entry and exit. Since greater uncertainty and sunk costs are associated with greaterprobability of bankruptcy and borrowing constraints, entry is likely to be affected for those firmsdependent on access to external funds. Similarly, to the extent that some incumbent firms are dependenton access to external credit, tightening credit conditions are likely to make their operations difficult, andin some circumstances may force them to exit. Overall, greater uncertainty and sunk costs may heightenasymmetric information problems and constrain borrowing, affecting both entry and exit.

2.2.(a). Links to Empirical AnalysisFinancing constraints appears to be a potentially important channel via which uncertainty and sunk costsmay affect entry, exit and industry structure. There is a fairly large literature that has come to identifysmall firms as the one’s most likely to be affected by information asymmetries and be financingconstrained; e.g., see Fazzari, Hubbard and Petersen (1988), Gertler and Gilchrist (1994), Ramey (1993),and the reference there. Gertler and Gilchrist note (p. 314):

“...while size per se may not be a direct determinant, it is strongly correlated with the primitivefactors that do matter. The informational frictions that add to the costs of external finance applymainly to younger firms, firms with a high degree of idiosyncratic risk, and firms that are notcollateralized. These are, on average, smaller firms.”13

We postulate that small firms are more likely to be affected by borrowing constraints under uncertainty.14

In addition, this effect is likely to be greater in industries with high sunk costs as lenders will be unableto recover much value from the collateral. The predictions can be summarized as:(A) Net entry. Greater uncertainty and sunk costs are likely to tighten borrowing constraints and delayentry. Exits may increase for firms more heavily dependent on borrowing; these are more likely to besmall firms. Thus, net entry is likely to be negative, with the industry most likely experiencing loss ofsmaller firms. In addition, these effects are likely to be greater in high sunk cost industries.(B) Size distribution. If small firms are more likely to be affected by borrowing constraints andexperience greater exit, the industry firm size distribution is likely to become less skewed with the effectbeing more pronounced in high sunk cost industries.(C) Concentration. Depends on the number of firms that exit and their share in total industry output.

Numerous papers examine the link between industry structure and innovation (see Cohen and Levin (1989)15

and Sutton (1997.b), and the reference there), but our focus is on the dynamics of technical change and industrystructure. The papers cited below provides detailed references to the related literature, which we do not repeat here.

Jovanovic and MacDonald argue that the Banbury mixer was the major innovation that caused a shakeout16

in the tire industry. But Klepper and Simons indicate it was the cord tire with its trajectory of innovations(continued...)

8

Given that smaller firms are more likely to be affected, the impact on concentration while positive, maynot be quantitatively large.

3. Technological Change and Other Factors

3.1. Technological ChangeWe focus on a class of models that link technological change to industry dynamics, and outline the basicresults. Gort and Klepper (1982) visualize two types of innovations: type I(1) constitutes ongoing15

improvements and emanates from incumbent firms, and type I(2) results from breakthroughs thattypically emanates from outside the industry to begin a new product cycle. The evolution consists of fivestages. As a result of I(2) innovation, the new product is introduced in Stage I with positive net entry,and Stage II is marked by greater positive net entry into the industry. Subsequent innovations are I(1)that lower costs and weed out inefficient firms: Stage III has net entry close to zero and Stage IV ischaracterized by high negative net entry, described as a “shakeout”. Gort and Klepper (p.634) write:

“[this] rise in innovation not only reinforces the barriers to entry but, in addition, compressesthe profit margins of the less efficient producers who are unable to imitate the leaders fromamong the existing firms. Consequently, the exit rate rises sharply until the less efficient firmsare forced out of the market.”

Stage IV leads into Stage V which continues until the eventual shrinkage of the market, induced byobsolescence of the product, or until fundamental new changes in technology launch a new productcycle. Their data on 46 industries provides evidence on the wide variation across industries in thepatterns of evolution as well as general confirmation of the link between patterns of technologicalchange and the net entry rates during Stages I-V.

Jovanovic and MacDonald (1994) and Klepper and Simons (2000) build on the above theme andprovide additional theoretical insights and evidence using data on the U.S. tire industry. These modelsassume a distribution of production efficiencies across firms, improvements in efficiency levels due tolearning-by-doing and imitation, and a low probability of successful innovations. Jovanovic andMacDonald consider the arrival of a “drastic” exogenous invention that decreases costs. The ensuingdynamics results in a shakeout with high negative net entry. A large number of the first-generation firmsexit, leaving behind bigger and more efficient survivors. In Klepper and Simons, each period gives riseto R&D opportunities to lower costs; innovators enjoy greater profit margins than imitators due to theirR&D activities. Due to increasing production efficiency, entry is reduced to a trickle and exit continueresulting in a shakeout. While Jovanovic and MacDonald rely on drastic innovations to explain industrydynamics, Klepper and Simons argue it may be a series of innovations culminating in a threshold withsignificant changes in net entry.16

(...continued)16

culminating in a breakthrough, the balloon tire. In a different context, Sutton (1997.b) provides numerous examplesof the varied patterns of technological change and entry and exit.

9

The above models assume convergence towards a steady state where the industry structurebecomes static. But Sutton (1997.a, p.52) notes that this is at odds with observed data which show highturnover of firms even in mature industries. Thus, even outside the framework of the above models,ongoing technological change is likely to play a key role in industry dynamics. In this respect, theempirical findings in Audretsch (1995) are interesting. Using a large sample of mature industries, hefinds that industry-wide innovation is negatively associated with new startups and survival of new firms,and smaller firms are more likely to exit. In contrast, new (small) firm survival rate is higher where newfirm innovation rate is higher. He attributes the negative impact of industry-wide innovation to Winter’s(1984) hypothesis of “routinized” regime where the stock of technology emanated from within theindustry and outsiders and new firms find it difficult to make inroads. The new firm innovation resultcorresponds to entrepreneurial, or non-routinized, regime which makes new entrants competitive.

3.1.(a). Links to Empirical AnalysisTechnological innovations occur continually and can be drastic or non-drastic; the latter are likely to bemore frequent. Incumbents and potential entrants have a distribution of production efficiencies. Basedon the above literature, an industry may experience negative net entry with each episode of non-drastictechnological change as some less efficient firms are weeded out. In contrast, drastic technical changemay result in positive net entry in the short-run as it may signal a new product cycle. Since industriesare likely to have experienced myriad forms of technical change over our sample period, in our empiricalanalysis we include a variable representing technological change (Section 4.4) and examine its impacton the number of (small and large) establishments and firms in an industry, and concentration.

3.2. Other FactorsThere are several other factors that will influence entry, exit and industry structure. Geroski (1995), thecollection of papers in Geroski and Schwalbach (1991) and Martin (1993, Ch.7), for example, outlinethese factors. In our empirical work we consider two variables in particular: industry growth -GROW-and profit margins -AA. The empirical evidence on the link between GROW and entry appears to bemixed. Audretsch (1995, p.61-63) finds that new startups are not affected by industry-specific growth,but are positively affected by broader macroeconomic growth. Data in Jovanovic and MacDonald (1994)indicate that the number of firms declined sharply when the industry was growing; also see Klepper(1999). As demonstrated in some of the empirical papers in Geroski and Schwalbach (1991) and thediscussion in Audretsch (Ch.3), the observed link between GROW and entry, exit and industry structureappears to be a bit tenuous. Thus, while GROW is an important control, its precise impact on industrystructure appears to be conditioned on the factors related to barriers to entry, macroeconomic conditionsand potentially the nature of transition the industry might be undergoing. Regarding profit-margin A,the expected sign in general is positive. But, as for GROW, it is likely to be conditioned on otherindustry-specific and aggregate factors. In the absence of entry barriers, greater A signals lucrativemarkets and may attract entry. However, some of the empirical results presented in the papers in Geroskiand Schwalbach indicate considerable variation in the coefficient of A. Geroski (1995) reviews someof the findings and notes that the reaction of entry to elevated profits typically appears to be slow. Thusthe empirical link between entry and profits appears to be a bit tenuous.

In section 4.2, following Sutton (1991), we construct a measure of minimum efficient scale (MES) for 1972,17

‘82 and ‘92 to proxy sunk entry capital costs. As noted there, the rank correlation between MES in 1972 and 1992is 0.94. A close look at the data summary statistics for the end-points indicated little change in the MES proxy.This provides some justification for arguing that industry fixed-effect may capture an important part of MES.

E.g., after controlling for industry fixed-effects, Domowitz et al (1987) find that advertising essentially has18

no effect on the time-series variations in industry markups. Their data shows far greater cross-industry variationin advertising than within-industry. They (p.25) “...conclude that by 1958, most of the industries in our samplehad already reached steady-state rates of advertising...” This provides an example of industry-fixed-effects (whichwe include in our model) capturing an important part of the impact of advertising. In addition, we note that theeffect of advertising on industry structure appears to be uncertain. Kessides (1986), e.g., shows that whileadvertising acts as a barrier to entry, it also increases the likelihood of an entrant’s success. In his study, the neteffect is that advertising has a positive effect on entry.

The FTC Line of Business data, e.g., on advertising and R&D are only available for about 4 years. Moreover,19

these data do not have a consistent level of disaggregation; some are at 3-digit and some at 4-digit level. Weexamined data on advertising available from the U.S. Statistics of Income: Corporate Source Book. These data aretypically at the 3-digit level, some only at the 2-digit level, and often there are important gaps which prevents usfrom constructing a consistent time series. Overall, the above mentioned data are not useful for our long time-seriesstudy. The industry studies I am aware of that include these variables are typically cross-section at a point in time,or over a very short time period. Our sample period is 1963-92 and I am not aware of any source that wouldprovide these data at our level of disaggregation. These deficiencies could be addressed in future studies that aremore industry-specific or as better data becomes available.

10

Our empirical analysis does not explicitly control for other variables; e.g., scale economies,advertising and R&D. This is due to the fact that we do not have time-series data in these variables overour sample period. We note the following. Scale economies are typically not likely to have large short-run fluctuations. If this is the case, then an industry-specific constant, which we include in our panel datamodel, will capture aspects of this relatively time-invariant component. Our empirical specification17

also includes a control for broadly defined technological change; one could argue that this variable maycapture aspects of scale economies. Finally, our empirical model includes a lagged dependent variable;to the extent that the lagged industry structure variable incorporates information on scale-economies inthe recent past, this provides an additional control. I am not aware of SIC 4-digit time-series data onadvertising or R&D for our sample of industries over 1963-92. To the extent that a part of advertisingand R&D intensities are in steady state levels and have a time-invariant component, this effect will bepartly captured by the industry-specific constant we include in our panel data model. Given that our18

empirical specification includes a time-series in broad technological change, this may partly captureR&D and related effects. Finally, since the lagged dependent industry structure variable capturesinformation on advertising and R&D from the recent past, it provides an additional control.19

4. Measurement

We use SIC 4-digit industry level data; Appendix A provides details. The industry-specific annual time-series data are over the period 1958-94. The industry structure data on the number of establishments,firms and concentration ratios, and data on sunk costs, were collected from the five-yearly Census ofManufactures; these data are not available at an annual frequency. This implies that in our empiricalestimation we use data at a five-yearly frequency. Below we describe the key variables used in our

As noted in section 2.1.(a), since our data do not contain information on entry and exit, we study net entry.20

11

empirical analysis, create the sunk costs sub-samples and present summary statistics.

4.1. Industry StructureWe collected data on the following industry structure variables: (i) total number of establishments in anindustry -ESTB; (ii) ESTB by size classes; (iii) the total number of firms in an industry -FIRMS; and(iv) the industry four-firm concentration ratio -CONC. A firm can have one or more establishments;20

an establishment is roughly defined as an economic entity operating at a location. The establishment sizeclasses are defined by the number of employees per establishment. The size classes available from theCensus are: number of employees 1-4; 5-9; 10-19; 20-49; 50-99; 100-249; 250-499; 500-999; 1,000-2,499; 2,500 or greater. The main focus of our empirical analysis will be on ESTB and by size classesas this allows us to examine the “size distribution”. Figures 2(a)-2(g) display the data on establishmentsize distribution over the seven Census years in our sample. The data show that typically about 25% ofthe total number of establishments in an industry belong to the smallest size category, and only about3% belong to the largest size group. The typical industry is thus heavily populated by smallestablishments. It is well known that larger (smaller) firms tend to be multi(single)-establishment. Giventhis, figures 2(a)-2(g) also roughly display the size distribution of firms. The data reveal a skewed sizedistribution of establishments for the typical industry as well as some fluctuations in this distributiontime. The skewed size distribution has been well documented in the literature (e.g., Simon and Bonini,1958; Ijiri and Simon, 1977; and Sutton 1997.a).

Next we use the ESTB data to construct samples of relatively small versus large establishments.The U.S. Small Business Administration (e.g., The State of Small Business: A Report of the President,1990) classifies a “small business” as one that employs less than 500 workers. This metric has been usedin public policy deliberations and lending policies towards small businesses. We use the 500 workercutoff as our benchmark: number of employees <500 constitutes our basic small business group, and$500 employees as the relatively large business group. Ghosal and Loungani (2000) provide additionaldiscussion of this benchmark. However, it could be argued that 500 employees may, in many instances,constitute a relatively large and wealthy business. Since there is no well defined scheme by which wecan define “small”, we create additional small business groups. Overall, our groups are: (i) Allindustries; (ii) relatively large businesses with $500 employees; (iii) relatively small businesses with<500 employees; and (iv) even smaller businesses as classified by (a) <250 employees, (b) <100employees and (c) <50 employees. In our empirical analysis, we examine the impact of uncertainty andsunk costs on ESTB and by size class. This may reveal whether uncertainty and sunk costs have adifferential impact on the small versus large establishments and may help us gain an understanding onhow they affect the size distribution. We also examine the impact on FIRMS and CONC.

Table 1 presents the within-year across-industry summary statistics on the industry structurevariables. The data indicate substantial variation across industries in the structure characteristics. Thedata also reveal the large share of small establishments for the typical industry. For the “representative”industry in our sample, there are (approximately) 640 ESTB, 588 FIRMS and CONC equal to 38%. Forthe typical industry, there are about 1.08 establishments per firm; the percentile distribution for the ratioof industry [#establishments/#firms] shows that the 50 , 75 and 90 values are approximately 1.1, 1.3th th th

and 1.6. The mean and median values imply near equivalence between the number of establishments andfirms in an industry; this is approximately true even at the 90 percentile value. This overall pictureth

conceals an important fact: larger (typically, older) firms generally tend to be multi-establishment (and

For example, the intensity of rental and used capital markets could be affected over booms and recessions.21

Shleifer and Vishny (1992) present models where asset resale value depends on industry-specific or economy-wideshocks. Using data that are long-run averages partly circumvents this problem.

12

multi-product), whereas smaller (typically, newer) firms are more likely to be single-establishment.For the purposes of our empirical analysis, it is important that there be “within-industry”

variation in the industry structure variables. Table 2 presents the within-industry summary statistics. Forexample, for the “representative” industry, the mean number of establishments is (approximately) 623and a within-industry standard deviation of 138; thus the coefficient of variation is about 22%. Lookingat the data in Table 2 on the number of establishments and by size classes, as well as data on FIRMS,we observe a reasonable amount of intertemporal within-industry variation. This is encouraging fromthe viewpoint of our proposed empirical examination of the impact of uncertainty on the dynamics ofindustry structure. Of the variables presented in Table 2, CONC has the lowest within-industry variation.

4.2. Sunk CostsFirst, we create measures of sunk costs. Second, we segment industries into low and high sunk costcategories. The theoretical models (Section 2) are primarily concerned with the fraction of initial entrycapital investment that is sunk; as this increases, the option value of waiting under uncertainty increases.These models treat sunk costs as proportional to entry capital requirements. While the theoretical issuesare relatively transparent, measurement of sunk cost presents difficult problems and there is littlesystematic empirical work that allows us to obtain good measures for a wide range of industries. Weadopt the methodology outlined in Kessides (1990) and Sutton (1991) to quantify sunk costs.

Drawing from the contestable markets literature, Kessides (1990) notes that the extent of sunkcapital outlays incurred by a potential entrant will be determined by the durability, specificity andmobility of capital. While these characteristics are unobservable, one can construct proxies. Let RENTdenote the fraction of total capital that a firm (entrant) can rent: RENT=(rental payments on plant andequipment/capital stock). Let USED denote the fraction of total capital expenditures that were on usedcapital goods: USED=(expenditures on used plant and equipment/total expenditures on new and usedplant and equipment). Finally, let DEPR denote the share of depreciation payments:DEPR=(depreciation payments/capital stock). High RENT implies that a greater fraction of capital canbe rented, implying lower sunk costs. Similarly, high USED signals an active market for used capitalgoods, implying lower sunk costs. High DEPR indicates that capital decays rapidly, implying lower sunkcosts - which arise from the undepreciated portion of capital. We collected data to construct RENT,USED and DEPR for the Census years 1972, 1982 and 1992. An advantage of using data at widelyspaced points in time is that they give us a picture of the long-run characteristics; the data are not likelyto be contaminated with short-run business cycle fluctuations.21

Next, we construct Sutton’s (1991) proxy for entry-capital sunk costs. In developing the theory(Sutton 1991, Ch.2), let M (>0) be defined as the setup cost or the minimal level of sunk cost that anentrant must incur, and S denote total industry sales (market size). Thus, in theory, M/S is the sunk costrelative to market size. In quantifying setup/sunk costs (Sutton 1991, Ch.4), he proposes a proxy thatmeasures the “relative” level of setup costs across industries. Sunk costs are assumed to be proportionalto the cost of constructing a single plant of minimum efficient scale (MES). Let S be a measure of MES,where S is the output of the median plant relative to industry output. Assume that the capital-sales ratioof the median firm is the same as the industry as a whole and denote industry capital-sales ratio by K/S.Then (M/S)=S(K/S). If we can obtain a proxy for S, and have data for industry K and S, we can

For example, (i) he assumes that the capital-output ratio of the median plant is representative of the entire22

industry, and this is unlikely to be the case; (ii) the book value of capital assets is used to compute the capital-salesratio, but the book value underestimates the current replacement cost; (iii) the computation assumes that the agestructure of capital does not vary across industries, and this is unrealistic.

Kessides (1990, Table 1), using data for 1982, reports summary statistics on USED, RENT and DEPR: the23

mean (s.d) values are 0.094 (0.100), 0.024 (0.025) and 0.070 (0.018). Our corresponding values (reported in Table3) averaged over 1972, 1982 and 1992 are roughly comparable.

13

approximate M/S. S is constructed using the distribution of plants within each SIC 4-digit industryaccording to employment size. The procedure used by Sutton (1991, p.98) to measure S is the oneimplemented by Kessides (1990). Let ‘m’ be the number of group sizes within the industry, and ‘n ’ andj

‘S ’ denote number of plants and total sales of the j size group (j=1,...,m.). Let Ms =(S /n );j j j jth

S =(1/m)G (Ms ); and S =G S . Then S=(S /S ). Using S and industry K/S, we obtain a proxy for M/S.e j j o j j e o

We label the term S(K/S) as SUNK(EC) (sunk costs-entry capital). As noted by Sutton (p.98), the cross-industry variation in SUNK(EC) provides a rough proxy for the cross-industry variation in sunk costs.Sutton notes many limitations of this measure. In addition, we note that SUNK(EC) is based on an22

estimate of the “median plant size”. As discussed in Section 2.1.(a), the typical entrant is very smallcompared to the typical incumbent, and it takes many years for new entrants to attain optimal scale. Thisimplies that the median plant size overstates the entry capital requirements of entrants. Further, this biasis likely to be larger in those industries where the optimal scale is relatively large, since the entrant islikely to be farther away from optimal scale. In this sense, sub-samples based on SUNK(EC) may be amore noisy indicator of the degree of sunk costs. Noting these limitations, we calculated SUNK(EC) forthe Census years, 1972, 1982 and 1992 (same time periods as for USED, RENT and DEPR).

Sunk Cost Sub-Samples: Table 3 presents summary statistics for RENT, USED, DEPR and SUNK(EC).The measures show a reasonable amount of cross-industry variation as evidenced by the value of thestandard deviation relative to the mean. We took a closer look at our measures for the end-points, 197223

and 1992. For the minimum efficient scale, MES, proxy S the rank correlation is 0.94, and 0.92 forSUNK(EC). This indicates a fair amount of stability in the MES and SUNK(EC) measures over oursample period. The mean (s.d.) for MES and SUNK(EC) were very similar over the end-points; the mean(s.d.) for USED, RENT and DEPR were relatively similar across time. We create high versus low sunkcost sub-samples by using the cross-industry median value. If SUNK(EC)<50 percentile, then sunkth

costs are low; high if SUNK(EC)$50 percentile. Similarly, sunk costs are low if RENT or USED orth

DEPR $50 percentile; high if RENT or USED or DEPR <50 percentile. In addition, we created a setth th

of sub-samples by combining the different characteristics. Low sunk costs if “USED and RENT andDEPR $50 percentile”; high if “RENT and USED and DEPR <50 percentile”. Finally, low sunk coststh th

if “USED and RENT and DEPR $50 and SUNK(EC) <50 percentile”; high if “RENT and USED andth th

DEPR <50 percentile and SUNK(EC) $50 percentile”. In combination, the alternate measures andth th

sub-samples serve as a check of robustness on our results related to uncertainty and sunk costs.

4.3. UncertaintyAs noted in Section 2, the stochastic element can be couched in terms of any relevant variable. We focuson a bottom line measure: uncertainty about profit-margins. Arguably, profit-margins are a key concernfor firms making entry and exit decisions. Commenting on the industry-specific determinants of the

This notion of uncertainty is consistent with previous work; see, e.g., Aizenman and Marion (1997), Ghosal24

(1996), Ghosal and Loungani (1996, 2000), Huizinga (1993) and Leahy and Whited (1996). These studies use thestandard deviation (or the conditional standard deviation) of some variable of interest as a measure of uncertainty.

This is consistent with theoretical definition of short-run profits (Varian, 1992, Ch.2). Empirically, this is a25

commonly used measure; see Carlton and Perloff (1994, Ch.9), Domowitz et al. (1986, 1987), Geroski and Mueller(1990), Ghosal (2000) and Machin and Van Reenen (1993). Carlton and Perloff (p.334-343) and Schmalensee(1989) discuss alternate measures and their pitfalls. Our measure A does not control for capital costs, which aremore important for measuring long-run profitability. As discussed by Carlton and Perloff and Schmalensee,quantifying capital costs is difficult due to problems related to valuing capital and assessing depreciation.

We present some summary statistics from the regressions -equation (3)- estimated to measure uncertainty.26

Across the 267 industries, the mean Adjusted-R and the standard deviation of adjusted-R were 0.62 and 0.25,2 2

respectively. The first-order serial correlation was typically low, with the cross-industry mean (std. dev acrossindustries) being -0.002 (0.07). Overall, the fit of the industry regressions was good.

14

turnover of firms, Sutton (1997.a, p.52-53), for example, notes the primary importance of the volatilityof industry profits. Dixit and Pindyck also discuss uncertainty about profits. We assume that firms usea profit forecasting equation to predict the level of future profits. The forecasting equation filters out thesystematic components. The standard deviation of the residuals (unsystematic component) are used asthe measure of the degree of profit-margin uncertainty. We measure industry profits by short-run profits24

per unit of sales. It is assumed that intermediate materials, energy and labor are variable inputs andcomprise the variable costs. Short-run profits are defined as:25

AA = [ (Total Sales Revenue minus Total Variable Costs)/(Total Sales Revenue) ],where variable costs include labor, materials and energy. The standard deviation of the unsystematiccomponent of A measures uncertainty. Later, in section 6, we discuss an alternate measure of profit-margins which accounts for depreciation expenses, and constructing the uncertainty measure from this.

4.3.(a). Benchmark Measure of UncertaintyThe profit forecasting equation includes lagged values of industry-specific sales growth (SALES) andeconomy-wide unemployment rate (UN). The justification for such a specification is contained in thestudies by Domowitz, Hubbard and Petersen (1986), Ghosal (2000) and Machin and VanReenen (1993)which show the sensitivity of profit-margins to industry-specific and aggregate economic conditions. Theprofit forecasting equation is given by (5), where A is the profit-margin of industry ‘i’ in time ‘t’:i,t

A = 8 + 8 A + 8 A + 8 SALES + 8 SALES + 8 UN + 8 UN + , (3).i,t 0 1 i,t-1 2 i,t-2 3 i,t-1 4 i,t-2 5 t-1 6 t-2 i,t

We use the following procedure to create a time-series for profit uncertainty:(i) For each industry in our sample, we first estimate equation (3) using annual data over the entiresample period 1958-1994. The residuals represent the unsystematic components.26

(ii) We use the standard deviation of residuals -FF(AA) - as our measure of uncertainty, where “i” and “t”i,t

index industry and time period. As noted earlier, industry structure data are for the years 1963, ‘67, ‘72,‘77, ‘82, ‘87 and ‘92. The standard deviation of residuals over, for example, 1967-71 serves as theuncertainty measure for 1972; similarly, s.d. of residuals over 1982-86 measures uncertainty for theCensus year 1987, and so on. Thus, we obtain seven time-series observations on F(A) .i,t

We also experimented with Autoregressive Conditional Heteroscedasticity (ARCH) models to measure27

uncertainty. After imposing the restrictions (Hamilton, 1994, Ch. 21), we estimated second-order ARCH for eachof the 267 industries. For about 45% of the industries the estimation failed to converge. Experimentation withalternate starting values, convergence criterion and order of the ARCH specification did not alleviate the basicproblems. This is probably not surprising given the limited number of time-series observations per industry. Wedid not pursue this approach any further.

15

Later we experiment with alternate specifications; our basic inferences are robust to these changes.27

Table 4 (column 1) presents the within-year cross-industry summary statistics for F(A). The standarddeviation is relatively high compared to the mean value indicating large cross-industry variation. Keyto our analysis, Table 5 (row 1) presents the “within-industry” summary statistics. For the representativeindustry, the ratio of the within-industry standard deviation (0.0117) to mean (0.0236) is about 50%,indicating a fair amount of within-industry variation in uncertainty. The range of within-industrystandard deviation is also large indicating wide variation in uncertainty across-industries. This variationin the data is encouraging for our empirical analysis.

4.3.(b). Alternate Measures of UncertaintyTo check the robustness of our inferences, we experimented with alternate specifications of the profitequation to construct the uncertainty measure. These included: (i) varying the lag length of theexplanatory variables in equation (3); (ii) following Ghosal (2000), replacing the broad business cycleindicator, unemployment rate, by the federal funds rate (FFR) and energy price growth (ENERGY); (iii)estimating a pure autoregressive model for the forecasting equation; and (iv) estimating the profitequation in growth rates instead of levels. We also experimented with an alternate measure of profit-margins that accounts for depreciation expenses. Our basic inferences do not change. In Section 6 wepresent additional details and results to confirm this.

4.4. TechnologyGiven our empirical framework, we need a time-series in technological change. Gort and Klepper(1982), Jovanovic and MacDonald (1994), Klepper and Simons (2000) and Klepper (1999) use specificinnovations for selected industries and attempt to categorize them as relatively drastic or non-drastic.However, these types of data are not available for our several hundred industries over our sample period.Cohen and Levin (1989) point to the considerable difficulties in measuring innovation. One set ofmeasures would be R&D or patents, but the required time-series data are not available for our industries.Further, Cohen and Levin, Audretsch (p.27-29) and several papers in Griliches (1984) highlight theproblems linking R&D expenditures to technological change and market opportunities, and theambiguities of patent data. Audretsch (1995) uses data on innovations commercially introduced in theU.S. in 1982, arguing that this is a better indicator of actual technological change than R&D and patents.But we don’t have a time-series for such data. Instead, we pursue an alternate strategy: we construct anindustry-specific time-series for technological change and use this as our control variable.

The Solow residual has been widely used to measure of technological change. Assuming (i)constant returns to scale, (ii) marginal cost pricing in product markets, (iii) marginal product basedpayments to factors and (iv) full and efficient utilization of inputs, the Solow residual -TECH(0)- froma four-factor (capital stock, K; labor hours H; materials, M; and energy, E) production function, is:

TECH(0) = [Îq - (( ªk +( ªh +( ªm +( ªe )] (4),t t kt t ht t mt t et t

Burnside et al. (1996) estimate (4) with annual data and find that TECH(1) has over 50% lower volatility28

compared to TECH(0), and much lower likelihood of technological regress, which is a desirable property.

Our inferences in section 6 are not affected if we use ªe instead of ªm. We also experimented with an29

alternate approach. In one of their specifications, Burnside et al. (1996) assume Leontief technology and grossoutput Q is produced with materials (M) and value-added (V): Q =min(" V , " M ), where "’s are constants. Vt V t M t

is produced according to a CRS production function using capital services (S) and labor hours (H): V=Z F(H ,t t t

S ), where Z is the exogenous technology shock. Capital services S is unobserved. In Burnside et al. (1996), S ist

assumed proportional to electricity consumption; Burnside (1996) assumes it is proportional to total energy usage.Since we do not have data on electricity usage, we proxy S by total energy usage (E): E= >S. Given this and theassumption of perfectly competitive factor markets, the factor utilization adjusted technology residual is:TECH(2)=[ªv - (1-" )ªh - " ªe ], where the lower-case letters denote logarithms of value-added, labor hourst Kt t Kt t

and energy (note that capital services is replaced by its proxy, energy usage). Using this approach to measure thetechnology residual did not alter our inferences.

16

where q, k, h, m and e are logarithms of output and the four inputs, and ( the input share.However, the literature has demonstrated that there can be significant deviation between true

technological change and TECH(0). One aspect that has received much attention is that inputs likecapital have variable “utilization” over the business cycle and this imparts a strong procyclical bias toTECH(0); see, e.g., Burnside et al. (1995), Burnside (1996) and Basu (1996). Various strategies havebeen adopted to correct TECH(0) of its cyclical input utilization bias: Burnside et al. (1995), followingJorgenson and Griliches’ (1967) suggestion, use electricity consumption to proxy utilization of capitaland obtain a corrected Solow residual; Burnside (1996) uses total energy consumption; and Basu (1996)uses materials inputs. The basic intuition is that these inputs (materials and energy) do not have acyclical utilization component like capital and, therefore, are good proxies for the utilization of capital;e.g., assuming capital stock is constant, if the utilization of capital increases, then materials and energyusage will typically increase. Collectively, these studies show that changes in factor utilization are animportant cause of the observed procyclical productivity and that changes in energy or materials usageare good proxies for the cyclical input utilization.

We construct a factor-utilization-adjusted technology residual following the insights in Burnside(1996) and Basu (1996). Burnside (1996) assumes that gross output Q is a differentiable function ofunobserved capital “services” (S), labor hours (H), materials (M) and energy (E): Q=Z F(S ,H ,M ,E ),t t t t t t

where Z represents exogenous technology shock. Assuming that S is proportional to materials usage(Basu, 1996), or energy consumption (Burnside, 1996), and competitive factor markets, the log-linearapproximation to the production function gives us the adjusted technology residual TECH(1):

TECH(1) = [ªq - (* ªe +* ªh +* ªm +* ªe )] (5),t Kt t Ht t Mt t Et t

where lower case letters denote logarithms, * is the share of the input in total revenue and ªs is replacedby ªm (as in Basu, 1996), or ªe (as in Burnside, 1996). In our experiments and empirical results, it did28

not matter whether we replaced ªs by ªm or ªe. Given that ªm is a broad based measure of input usage,we report results with ªm as the proxy for ªs. We use TECH(1) as our benchmark measure of29

technological change. Table 4 (column 2) presents the within-year cross-industry summary statistics onTECH(1). Table 5 (row 2) presents the within-industry summary statistics. These data indicate highcross-industry as well as within-industry variation in the rate of technological change.

In theory, an entrant should rationally expect profit-margins to fall post-entry, implying that we construct30

expected post-entry margins. However, our typical industry contains about 560 firms (see Tables 1 and 2). Giventhis large base of incumbents, in general it appears unlikely that an increment of “one” entrant would affect pricesand margins. Further, in section 2.1(a) we noted that the typical entrant is very small compared to the incumbentsand assume a minuscule share of the market. Given the size of the entrants, it again appears very unlikely thatthey’ll have an impact on industry prices and margins. Given these considerations, we do not attempt to constructmeasures of expected post-entry margins. Our approach implies that entrants assume that pre-entry profit-marginswill prevail post-entry, and this is meaningful given the entrants’ size and the large number of incumbents.

Martin (1993, Ch.7), for example, reviews some studies that have used similar models.31

17

4.5. Other VariablesThe final two variables are industry profit-margins -AA- and growth -GROW. A is measured as describedin section 4.3. The industry structure variable in period ‘t’ is explained by A over the preceding period;e.g., the number of firms in 1972 is explained by the mean level of A over 1967-1971. Apart from30

using the mean level of A, we also experimented with using the growth rate of A over the precedingperiod. Our basic inferences did not change. Table 4 (column 3) and Table 5 (row 3) present the cross-industry and within-industry summary statistics on A. Our proxy for industry growth is the mean rateof new investment. New investment entails sunk costs; thus if new (net) investment is increasing, it islikely to indicate expanding market opportunities. As is standard (Fazzari et al., 1988), we measure netinvestment by the ratio (I /K ), where I is total industry investment in the current period and K isi,t i,t-1 i,t i,t-1

the end of last period capital stock. The industry structure variable in period ‘t’ is explained by the meanrate of net investment over the preceding period; e.g., the number of firms in 1972 is explained by themean rate of net investment over 1967-1971. Table 4 (column 3) and Table 5 (row 3) present the cross-industry and within-industry summary statistics on GROW. As a check of robustness for our uncertaintyresults, in Section 6 we also report estimates using industry sales growth as a proxy for growth; ourresults regarding uncertainty are not affected.

5. Panel Data Model

We use a dynamic panel data model to examine the impact of uncertainty, sunk costs and technology.Entry and exit are not likely to occur instantaneously to restore an industry’s equilibrium under changingconditions. Furthermore, there is uncertainty regarding the time is takes to restore equilibrium. Giventhese considerations, we use a partial adjustment model to structure our empirical equation. Denoting31

industry structure by STR, where STR could be ESTB (and by size groups), FIRMS or CONC, we get:

STR = 8STR - (1-8)STR (6),i,t i,t i,t-1*

where ‘i’ and ‘t’ denote industry and time, STR the equilibrium structure in period t, and 8 the partial-*

adjustment parameter. STR is not observed and is modeled as a function of the following industry-*

specific variables: (i) industry-specific profit uncertainty, F(A) ; (ii) industry-specific technologicali,t

change, TECH(1) ; (iii) industry-specific profit-margin, A ; and (iv) industry-specific growth, GROW .i,t i,t i,t

Apart from (i)-(iv), the panel data model includes the following additional controls: (v) an industry-specific fixed-effect " to control for unobserved factors that influence the long-run level of industryi

structure. These would be unobserved relatively time-invariant elements of scale economies, andadvertising and R&D intensities (see our discussion in section 3.2); and (vi) an aggregate structure

Hausman tests (see notes to Table 7) easily reject the null that the industry variables are pre-determined.32

18

variable, ASTR, to control for manufacturing-wide effects that are common to all industries. Audretsch(1995, Ch.3), for example, finds that macroeconomic factors play an important role in determining entry;variations in ASTR will capture these aggregate effects.

Incorporating these features, the dynamic panel data model is given by:

STR = " + > F(A) + > TECH(1) + > A + > GROW + > ASTR + > STR + , (7).i,t i 1 i,t 2 i,t 3 i,t 4 i,t 5 t 6 i,t-1 i,t

In equation (7), the variables STR, F(A), A, GROW and ASTR are measured in logarithms; thus thesecoefficients can be interpreted as elasticities. TECH(1) is not measured in logarithms as it can (and does)take on both negative and positive values (see section 4.4 for construction of TECH). Next, to clarifythe variables in (7), suppose STR is FIRMS . Then F(A) is the standard deviation of residualsi,t i,1972 i,1972

over 1967-1971; TECH(1) is the mean rate of technical change over 1967-71; A is the meani,1972 i,1972

profit-margin over 1967-71; GROW is the mean rate of net investment over 1967-71; AFIRMSi,1972 1972

is the total number of firms in manufacturing in 1972; and FIRMS (the lagged dependent variable)i,1967

is the total number of firms in the 4-digit industry in 1967. As discussed in section 3.2, the laggeddependent industry structure variable will capture aspects of scale economies, and advertising and R&Dintensities using information from the recent past. We estimate (7) for all industries in our sample as wellas the sunk cost sub-samples. Table 6 presents the summary statistics from our panel data.

Estimation MethodFirst, as shown in the literature on estimation of dynamic panel data models, we instrument STR , thei,t-1

lagged dependent variable. Second, industry-specific variables like the number of establishments andfirms, profit-margins, output, input usage, technological change (constructed from data on industryoutput and inputs) are all likely to be jointly-determined in industry equilibrium. Thus, the industry-specific variables are best treated as endogenous. Several estimators have been proposed to obtain32

efficient and unbiased estimates in dynamic panel models; see, e.g., Baltagi (1994, Ch.8), Kiviet (1995),and the reference there. Our estimation proceeds in two steps. First, we sweep out the industry intercept" by taking deviations from within-industry means; the data are now purged of systematic differencesi

across industries in the level of the relevant structure variable.Second, the within-industry equation is estimated using the instrumental variable (IV) estimator,

treating F(A) , TECH(1) , A , GROW , and STR as endogenous. We include a broad set ofi,t i,t i,t i,t i,t-1

instruments as the literature indicates this is needed to alleviate problems related to bias and efficiency.The variables and their instruments are:(i) F(A) : Instrumented by F(A) and F(A) . While we include lagged instruments, since our data arei,t i,t-1 i,t-2

over 5-year time intervals (e.g., F(A) is constructed using data over 1972-1976), we also includei,1977

instruments constructed at a higher level of aggregation that would be correlated with F(A) andi,t

uncorrelated with the error term. The objective is to provide a stronger overall set of instruments. Weadopt the following procedure: we segment our data into durable (D) and non-durable (ND) goodsindustries. The business cycle literature indicates that these two categories of industries show markedlydifferent patterns of fluctuations. There are 450 SIC 4-digit industries in the manufacturing sector andthese are roughly evenly split between D and ND industries. It is unlikely that any one D or ND 4-digitindustry will be influencing all the D or ND industries. Fluctuations in the entire group of D or NDindustries will be driven by factors exogenous to a given industry; thus constructing instruments at the

19

D/ND level appears to be reasonable. The instrument that we construct for F(A) is the standardi,t

deviation of D/ND profit-margins over the relevant period. For example, for F(A) the instrumenti,1977

is the standard deviation of A (for D and ND) over 1972-1976: we label this as FF(AA: D/ND) .t(ii) TECH(1) : Instrumented by TECH(1) and TECH(1) . As with the uncertainty variable above,i,t i,t-1 i,t-2

we include TECH constructed at the D/ND level - TECH(1: D/ND) - as an instrument.t

(iii) A : Instrumented by A and A . In addition, we include A constructed at the D/ND level -i,t i,t-1 i,t-2

AA(D/ND) - as an instrument.t

(iv) GROW : Instrumented by GROW and GROW . We also include the D/ND growth variable -i,t i,t-1 i,t-2

GROW(D/ND) .t(v) STR Instrumented by STR and the manufacturing-wide ASTR and ASTR since ASTR isi,t-1: i,t-2 t t-1

included in equation (7) and can be treated as exogenous to a given 4-digit industry.In section 6 we carry out several checks to gauge the robustness of our uncertainty results.

6. Estimation Results

First, we discuss the findings from the full sample of industries. Second, we examine the results fromthe sunk-cost sub-samples. Third, we present some additional results to check the robustness of ourresults related to uncertainty and sunk costs.

Full SampleTable 7 presents the results from estimating equation (7). First, we focus on the F(A) estimates; thesecoefficients are interpreted as elasticities since the variables are measured in logarithms. The F(A)coefficients are negative and significant for all establishments and the small establishment groups; theF(A) coefficient is positive and insignificant for the large establishment group. As the establishment sizegets smaller (e.g., Size<500; Size<250; Size<100; and Size<50), the F(A) elasticity gets noticeablylarger. Thus, greater uncertainty primarily impacts small establishments. Turning to the other industrystructure variables, greater F(A) decreases FIRMS. The summary statistics presented in Section 4.1indicate a rough equivalence between an establishment and a firm; the 50 (75 ) percentile value of theth th

ratio [#establishments/#firms] was 1.1 (1.3). Further, it is well known that larger (smaller) firms typicallytend to be multi(single)-establishment. Given this, we conclude there is a reduction in the number ofsmall firms in an industry. Finally, the F(A) elasticity is positive for CONC. In the table below wepresent the quantitative effect of a one-standard-deviation increase in F(A).

Quantitative effect of a one-standard-deviation increase in F(A).

Number of Establishments by Size Group

Size: All Size: $500 Size: <500 Size: <250 Size: <100 Size: <50 FIRMS CONC

-72 +0.7 -74 -84 -100 -103 -60 +5

The numbers indicate, for example, that a one-s.d. increase in F(A) results in a drop of 60 firms over the5-year Census interval, starting from a mean value of 558 firms. Similarly, a one-s.d. increase in F(A)results in about 5 point increase in the four-firm concentration ratio, starting from a mean value of about39%. These represent meaningful economic effects. Overall, uncertainty reduces the number of smallfirms and establishments, and increases industry concentration; given the results on small v. large

Our result that uncertainty has an adverse impact on the number of small firms and establishments resonates33

with the findings in Ghosal and Loungani (2000) where greater uncertainty decreased investment spending bysmaller firms, with large firm investment typically remaining unaffected.

20

establishments (firms), we can say that the establishment (firm) size distribution becomes less skewed.33

Next we examine the effect of technological change, noting that TECH(1) is not measured inlogarithms. Technological improvement reduces the number of small establishments; the impact on largeestablishments is positive but insignificant. The point estimate of TECH(1) gets noticeable larger in thesmaller establishment groups. Given the correspondence between small establishments and firms, greaterTECH(1) reduces the number of small firms. The impact on industry CONC is positive, but thecoefficient is not statistically significant. As for F(A), we take a look at the quantitative effects. The tablebelow shows the effect of a one-standard-deviation increase in TECH(1).

Quantitative effect of a one-standard-deviation increase in TECH(1).

Number of Establishments by Size Group

Size: All Size: $500 Size: <500 Size: <250 Size: <100 Size: <50 FIRMS CONC

-26 +0.1 -26 -32 -38 -40 -22 +1.6

For example, that a one-s.d. increase in TECH(1) results in a decrease of 40 firms over the 5-year Censusinterval, starting from a mean value of 558 firms. Similarly, a one-s.d. increase in TECH(1) results inabout 1.6 point increase in CONC, starting from a mean value of about 39%. As with the uncertaintyeffects, these represent meaningful economic effects. However, from our estimated regressions, thequantitative effect of uncertainty is greater than technological change. Overall, technological changereduces the number of small firms and establishments, increases industry concentration and appears tomake the establishment (firm) size distribution becomes less skewed.

Industry profit-margins, A, appear to have no effect on the number of small establishments andfirms, or in the full sample. The number of large establishments increase following an increase in profit-margins; industry CONC rises. Industry growth, GROW, has a negative and significant effect on thenumber of large establishments, and a weak negative effect in the full sample. The general ambiguityof the profit and growth results appear to be similar to those observed in the previous literature (see ourdiscussion in section 3.2). The industry structure variables in general co-vary positively with theiraggregate (ASTR) counterparts; the exceptions being the number of large establishments. Finally, apartfrom CONC, all the lagged industry structure variables are positive and significant.

Sunk Cost Sub-SamplesSince the focus of our study is on the impact of uncertainty and sunk costs, in Table 8 we only presentthe estimates of F(A). For ease of comparison, the first column reproduces the full-sample estimatesfrom Table 7. The following observations emerge:(i) For all establishments (row: Size All), greater F(A) has a statistically significant negative effect onlyin the high sunk costs sub-samples; while the elasticities vary somewhat across the alternate sunk costsmeasures, the qualitative inferences are similar. The F(A) elasticities are insignificant in the low sunkcost sub-samples;

21

(ii) For the large establishment group (row: Size $500), the F(A) estimates are statistically insignificantand typically positive; the only exception being the DEPR high sunk cost sub-sample where the F(A)coefficient is negative and significant.(iii) Greater F(A) reduces the number of small establishments only in the high sunk cost groups. And,as the size class get smaller (Size<500; ...; Size<50), the F(A) elasticities get larger in the high sunk costcategories. The exception being the SUNK(EC) groups where greater F(A) reduces the number of smallestablishments even when sunk cost are low, but the elasticities are larger in the high sunk cost group.(iv) For FIRMS, the F(A) elasticities are negative and significant only in the high sunk cost sub-samples.The only close call is for the low SUNK(EC) group where the elasticity is negative and close tosignificance at the 10% level. Given the rough equivalence between the number of establishments andfirms, and the results in (iii), uncertainty reduces the number of small firms in high sunk cost industries.(v) The F(A) elasticities are positive for the CONC regressions, but they are statistically significant onlyin the high sunk cost sub-samples.

In Table 9 we present results from some alternate sunk cost sub-samples. As described in section4.2, these sub-samples are created by combining the alternate measures. For comparison, the firstcolumn reproduces the results from Table 7. As in Table 8, we find that greater uncertainty reduces thenumber of small establishments and firms, and increases concentration, in the high sunk cost industries.The elasticities presented in Table 9 present a much starker effect of uncertainty on small firm dynamics.As before, uncertainty does not appear to have an effect on the number of large establishments in anindustry irrespective of the degree of sunk costs. The broad picture that emerges from Table 8 and Table9 can be summarized as follows. In general, greater uncertainty reduces the number of smallestablishments and firms only in the high sunk cost sub-samples; the impact is typically insignificant inthe low sunk cost groups. Greater uncertainty typically has no impact on the number of largeestablishments. The effects on small v. large establishments (firms) implies that greater uncertainty leadsto a less skewed distribution of establishment (firm) sizes only in the high sunk cost industries.

Some Checks of RobustnessTo gauge the robustness of our uncertainty results, we carried out several checks. Some of these resultsare reported in Table 10. Since the focus of this paper is on the effect of uncertainty and sunk costs, weonly report the F(A) estimates. Panel A reproduces the estimates from Table 7 for reference.(A) We experimented with alternate specifications of the profit forecasting equation to construct theuncertainty measure. First, following Ghosal (2000), we replace the broad business cycle indicator,unemployment rate, by the federal funds rate (FFR) and energy price growth (ENERGY) and constructthe uncertainty measure from these residuals. These results appear in panel B. Second, we estimated anAR(2) specification; these results appear in panel C.(B) We constructed an alternate measure of industry profit-margins by accounting for depreciationexpenses. The data on industry-specific depreciation rates were collected for the Census years 1972,1982 and 1992 (same as those used to create the DEPR sub-samples). We assumed that these data arerepresentative for the entire sample period and constructed the alternate measure as:AA(alt)=[(Total Sales Revenue-Total Variable Costs-Depreciation Expenses)/(Total Sales Revenue)].Using this measure, we reestimated equation (3) to construct F[A(alt)]. We did not report these as ourmain results since we do not have a time-series in depreciation rates which would be required to makea proper comparison with our main measures A and F(A). The results using F[A(alt)] are in panel D.(C) In our main specification, we used the mean rate of new investment as our proxy for industry growthconditions, GROW. We experimented with an alternate measure, the growth of industry sales, and re-

22

estimated equation (7). These results appear in panel E.While in Table 10 we only report the equivalent of Table 7 estimates, we also examined the equivalentof Tables 8 and 9 sunk cost sub-sample estimates; we do not present the latter as they would be veryspace consuming. While there were some quantitative differences, the experiments A-C above did notalter our basic inferences from Tables 7-9.

We also experimented with the following:(i) Varying the lag length of the explanatory variables in equation (3);(ii) Estimating the profit equation in growth rates instead of levels; and(iii) An alternate instrument for F(A) by constructing the (durable/non-durable) D/ND profit uncertaintyinstrument (see section 5) by estimating a forecasting equation and using the residuals, instead of simplytaking the standard deviation of D/ND profits. We used annual (1958-94) data on D and ND profit-margins and separately estimated: A =J +J A +J A +J SALES +J SALES +J UN +J UN +, .t 0 1 t-1 2 t-2 3 t-1 4 t-2 5 t-1 6 t-2 t

The uncertainty measure is constructed using the standard deviation of residuals.None of these experiments altered our basic inferences reported in Tables 7-9.

7. Discussion and Concluding Remarks

The primary focus of our study was on uncertainty and sunk costs and our results indicate that they maybe important influences affecting firm dynamics and industry structure. The empirical findings in thispaper lend support to Sutton’s (1997.a, p.53) insight that fluctuations in industry-specific profits mightbe of primary importance in understanding firm dynamics. Our study reveals the following: (i) greateruncertainty has a negative impact on the number of establishments and firms, and a positive impact onconcentration; (ii) the negative impact is on small establishments and firms, with large firms typicallybeing unaffected; (iii) the negative impact on the number of small establishments and firms appearsmainly in the high sunk cost sub-samples; and (iv) results (ii) and (iii) imply that uncertainty and sunkcosts result in a less skewed distribution of establishment (firm) sizes in the high sunk cost groups. Howcan we reconcile these findings with respect to the “option value” and “asymmetric information”channels discussed in section 2? With respect to the option value channel, we could not make clearpredictions whether the number of firms and establishments should increase or decrease under greateruncertainty and sunk costs. Neither could we make clear predictions regarding the small versus largefirm issue. If it were true that uncertainty and sunk costs played a greater role in reducing entry leavingexits relatively unaffected, and this was especially the case for smaller firms, then our empirical findingswould be consistent with this result. In contrast, the asymmetric information channel provided relativelyclearer insights. In particular, we noted that smaller firms were more likely to be adversely affected byinformation asymmetries in capital markets than larger firms. The implication being that greateruncertainty and sunk costs, which exacerbated the information asymmetries, would adversely affectsmaller firms (likely entrants and incumbents), potentially lowering their entry, increasing exits andresulting in a reduction in the number of establishments and firms in an industry. While we were not ableto precisely disentangle the option value versus asymmetric information channels, our results appear tolend more support to the latter.

A greater rate of technological change appears to reduce the number of small firms andestablishments, with little effect on the number of larger firms and establishments. Although we use avery different measure of technological change (a variant of the Solow residual) than employed in theprevious literature (e.g., R&D, innovations, patents) on firm dynamics, our findings appear to be similarin spirit to those in Audretsch (1995) who finds that a greater rate of industry-wide innovation adversely

23

affects new startups, favoring incumbents. Audretsch noted that his findings were consistent with thehypothesis of a routinized technological regime (Winter, 1984). However, our findings also appear tobe consistent with, for example, some of the notions outlined in Klepper and Simons (2000), amongothers, where incremental technological change decreases costs and forces out inefficient firms. Sincethe role of technology was not central to our study, we did make an attempt to differentiate across thealternate channels via which technology affects firm dynamics. This is left to future research.

Our findings on uncertainty and sunk costs could have implications for research in several areas.First, an important part of competition policy analysis is to assess factors related to the height of entrybarriers, likelihood of entry and evaluate other forces that govern competition in markets. Based on ourresults, it appears that uncertainty in combination with sunk costs may act as a natural barrier to entryand potentially affect competition via reduction in the number of industry participants and increasedconcentration. Second, high job creation and destruction rates are well documented (e.g., Davis,Haltiwanger and Schuh, 1996). Third, the high intertemporal volatility of investment spending is wellknown. If entry and exit reflect the bigger picture of economic activity in industries, then our resultscould imply that greater uncertainty and sunk costs influence job turnover and investment spending.Fourth, Caballero and Hammour (1994), for example, study the process of creative destruction inresponse to economic fluctuations. Our results on firm turnover induced by uncertainty and sunk costscould shed additional light on these models. Fifth, since our results suggest that uncertainty and sunkcosts affect the firm size distribution, future research could explore the role played by additionalchannels such as those suggested by Kihlstrom and Laffont (1979) and Lucas (1978). Sixth, insightsfrom our analysis could be applied to events in specific industries. For example, the electric industry isundergoing deregulation. In recent years this industry has shown a high degree of uncertainty aboutprofits resulting in many firms experiencing financial difficulty. Finally, the industry is going througha spate of mergers involving firms of different sizes. While a proper analysis of the evolution of thisindustry would require a detailed study, our findings could be used to predict a future path for thisindustry that leads to greater concentration and weeding out of smaller firms. More generally, while thefocus of this paper was on an inter-industry study to get broad insights, industry-specific studies that maybetter control for underlying factors related to sunk costs, uncertainty and technological change mayyield greater dividends. Finally, constrained by the data, we could only examine patterns of net entry(establishments and firms). Availability of long time-series data on entry and exit would allow us todisentangle the effects of uncertainty and sunk costs on these variables and may provide deeper insights.

24

References

Appelbaum, Elie, and Eliakim Katz. “Measures of Risk Aversion and Comparative Statics of IndustryEquilibrium,” American Economic Review 76 (1986), 524-29.

Appelbaum, Elie, and Chin Lim. “Contestable Markets under Uncertainty,” RAND Journal of Economics16, 1985, 28-40.

Audretsch, David. Innovation and Industry Evolution. Cambridge: MIT Press, 1995.

Baldwin, Richard and Paul Krugman. “Persistent Trade Effects of Large Exchange Rate Shocks,”Quarterly Journal of Economics 104, 1989, 635-654.

Baltagi, Badi. Econometric Analysis of Panel Data. New York: John Wiley, 1995.

Bartlesman, Eric, and Wayne Gray. “The Manufacturing Industry Productivity Database,” NationalBureau of Economic Research, 1998.

Basu, Susanto. “Procyclical Productivity: Increasing Returns or Cyclical Utilization?” Quarterly Journalof Economics 111, 1996, 719-751.

Baumol, William, John Panzar, and Robert Willig. Contestable Markets and the Theory of IndustryStructure. San Diego: Harcourt Brace Jovanovich, 1982.

Burnside, Craig, Martin Eichenbaum and Sergio Rebelo. “Sectoral Solow Residuals,” EuropeanEconomic Review 40, 1996, 861-869.

Burnside, Craig, Martin Eichenbaum and Sergio Rebelo. “Capital Utilization and Returns to Scale,”National Bureau of Economic Research Macroeconomics Annual, 1995, 67-119.

Caballero, Ricardo, and Mohamad Hammour. “The Cleansing Effect of Recessions,” AmericanEconomic Review 84, 1994, 1350-1368.

Chan, K.C., and Nai-Fu Chen. “Structural and Return Characteristics of Small and Large Firms,”Journal of Finance 46, 1991, 1467-1484.

Cohen, Wesley and Richard Levin. “Empirical Studies of Innovation and Market Structure” inSchmalensee, Richard and Robert Willig, ed., Handbook of Industrial Organization, Amsterdam: NorthHolland, (1989).

Davis, Steven, John Haltiwanger and Scott Schuh. Job Creation and Destruction. Cambridge: MIT Press,1996.

Dixit, Avinash. “Hysteresis, Import Penetration, and Exchange Rate Pass-Through,” Quarterly Journalof Economics 104, 1989, 205-28.

25

Dixit, Avinash. “Entry and Exit Decisions under Uncertainty,” Journal of Political Economy 97, 1989,620-638.

Dixit, Avinash. “Analytic Approximations in Models of Hysteresis,” Review of Economic Studies 58,1991, 141-151.

Dixit, Avinash, and Robert Pindyck. Investment under Uncertainty. Princeton: Princeton UniversityPress, 1996.

Domowitz, Ian, Glenn Hubbard and Bruce Petersen. “Business Cycles and the Relationship BetweenConcentration and Price-Cost Margins,” RAND Journal of Economics 17, 1986, 1-17.

Domowitz, Ian, Glenn Hubbard, and Bruce Petersen. “Oligopoly Supergames: Some Evidence on Pricesand Margins,” Journal of Industrial Economics 35, 1987, 379-98.

Dunne, Timothy, Mark Roberts, and Larry Samuelson. “Patterns of entry and exit in U.S. manufacturingindustries,” RAND Journal of Economics 19, 1988, 495-515.

Evans, David. “The Relationship Between Firm Growth, Size and Age: Estimates for 100 ManufacturingIndustries,” Journal of Industrial Economics 35, 1987, 567-581.

Fazzari, Steven, Glenn Hubbard and Bruce Petersen, “Financing Constraints and Corporate Investment.”Brookings Papers on Economic Activity 1 (1988), 141-195.

Gale, Douglas, and Martin Hellwig. “Incentive-Compatible Debt Contracts: The One Period Problem,”Review of Economic Studies 52, 1985, 647-663.

Geroski, Paul. “What do we know about Entry?” International Journal of Industrial Organization 13,1995, 421-440.

Geroski, Paul, and Joachim Schwalbach. Entry and Market Contestability. Oxford: Blackwell, 1991.

Gertler, Mark, and Simon Gilchrist. “Monetary Policy, Business Cycles, and the Behavior of SmallManufacturing Firms,” Quarterly Journal of Economics 108 (1994), 309-340.

Ghosal, Vivek. “Product Market Competition and Industry Price-Cost Margin Fluctuations: Role ofEnergy Price and Monetary Changes,” International Journal of Industrial Organization 18, 2000, 415-444.

Ghosal, Vivek. “Does Uncertainty Influence the Number of Firms in an Industry?” Economics Letters50, 1996, 229-236.

Ghosal, Vivek and Prakash Loungani. “The Differential Impact of Uncertainty on Investment in Smalland Large Businesses,” The Review of Economics and Statistics 82, 2000, 338-343.

26

Ghosal, Vivek and Prakash Loungani. “Product Market Competition and the Impact of Price Uncertaintyon Investment,” Journal of Industrial Economics 44, 1996, 217-228.

Gort, Michael and Steven Klepper. “Time Paths in the Diffusion of Product Innovations,” EconomicJournal 92, 1982, 630-653.

Greenwald, Bruce, Joseph Stiglitz, and Andrew Weiss. “Information Imperfections in the Capital Marketand Macroeconomic Fluctuations,”American Economic Review, 1984, 194-199.

Greenwald, Bruce and Joseph Stiglitz. “Macroeconomic Models with Equity and Credit Rationing,” inHubbard, R. Glenn. ed., Asymmetric Information, Corporate Finance, and Investment. Chicago:University of Chicago Press, 1990, 15-42.

Griliches, Zvi (ed). R&D, Patents and Productivity. Chicago: University of Chicago Press, 1984.

Hopenhayn, Hugo. “Entry, Exit and Firm Dynamics in Long Run Equilibrium,” Econometrica 60, 1992,1127-1150.

Ijiri,Yuji, and Herbert Simon. Skew Distributions and Sizes of Business Firms. Amsterdam: NorthHolland, 1977.

Jovanovic, Boyan and Glenn MacDonald. “The Life Cycle of a Competitive Industry,” Journal ofPolitical Economy 102, 1994, 322-347.

Kessides, Ioannis. “Market Concentration, Contestability and Sunk Costs,” Review of Economics andStatistics 72, 1990, 614-622.

Kessides, Ioannis. “Advertising, Sunk Costs and Barriers to Entry,” Review of Economics and Statistics68, 1986, 84-95.

Kihlstrom, Richard and Jean-Jacques Laffont. A General Equilibrium Entrepreneurial Theory of FirmFormation Based on Risk Aversion,” Journal of Political Economy 87, 1979, 719-748.

Kiviet, Jan. “On Bias, Inconsistency, and the Efficiency of Various Estimators in Dynamic Panel DataModels,” Journal of Econometrics 68, 1995, 53-78.

Klepper, Steven and Kenneth Simons. “The Making of an Oligopoly: Firm Survival and TechnologicalChange in the Evolution of the U.S. Tire Industry,” Journal of Political Economy 108, 2000, 728-760.

Klepper, Steven. “Firm Survival and Oligopoly,” Mimeo, Carnegie Mellon University, 1999.

Lambson, Val. “Industry Evolution with Sunk Costs and Uncertain Market Conditions,” InternationalJournal of Industrial Organization 9, 1991, 171-196.

Lucas, Robert. On the Size Distribution of Business Firms,” Bell Journal of Economics, 1978, 508-523.

27

Machin, Stephen, and John Van Reenen, “Profit Margins and the Business Cycle: Evidence from U.K.Manufacturing Firms,” Journal of Industrial Economics 42 (1993), 29-50.

Martin, Stephen. Advanced Industrial Economics. Oxford: Blackwell, 1993.

Myers, S., and N. Majluf. “Corporate Financing Decisions when Firms have Investment Information thatInvestors do not,” Journal of Financial Economics 13, 1984, 187-220.

Pakes, Ariel, and Richard Ericson. “Empirical Implications of Alternate Models of Firm Dynamics,”Journal of Economic Theory 79, 1998, 1-45.

Ramey, Valerie. “How Important is the Credit Channel in the Transmission of Monetary Policy?”Carnegie-Rochester Conference Series on Public Policy, 1993, 1-46.

Schmalensee, Richard, “Inter-Industry Studies of Structure and Performance,” in Schmalensee, Richardand Robert Willig, ed., Handbook of Industrial Organization, Amsterdam: North Holland, (1989).

Shleifer, Andrei, and Robert Vishny. “Liquidation Values and Debt Capacity: A Market EquilibriumApproach,”Journal of Finance 47, 1992, 1343-1366.

Simon, Herbert, and Charles Bonini. “The Size Distribution of Business Firms,” American EconomicReview 48, 1958, 607-17.

Spencer, Barbara, and James Brander. “Pre-Commitment and Flexibility: Applications to OligopolyTheory,” European Economic Review 36, 1992, 1601-1626.

Stiglitz, Joseph, and Andrew Weiss. “Credit Rationing in Markets with Imperfect Information,”American Economic Review, 1981, 393-410.

Sutton, John. Sunk Costs and Market Structure. Cambridge: MIT Press, 1991.

Sutton, John. “Gibrat’s Legacy,” Journal of Economic Literature 35, 1997(a), 40-59.

Sutton, John. Technology and Market Structure. Cambridge: MIT Press, 1997(b).

Williamson, Oliver. “Corporate Finance and Corporate Governance,” Journal of Finance 43, 1988, 567-591.

Winter, Sidney. “Schumpeterian Competition in Alternative Technological Regimes,” Journal ofEconomic Behavior and Organization 5, 1984, 287-320.

28

Appendix A: Data

The table below summarizes the data sources and years for which they are available. The industry dataare at the SIC 4-digit level of disaggregation. The following industries were excluded from the sample:(i) “Not elsewhere classified” since they do not correspond to well defined product markets;(ii) Industries that could not be matched properly over time due to SIC definitional changes; there wereimportant definition changes in 1972 and 1987. For these industries, the industry time-series and otherstructural characteristics data are not comparable over the sample period; and(iii) Industries that had missing data on the industry structure and sunk cost variables.The final sample contains 267 SIC 4-digit manufacturing industries. Given the above exclusions, thefinal sample contains a balanced panel of industries that are well defined over the sample period andhave data consistency.

Variable Source Years Available

Industry time-series data: Bartlesman and Gray (1998). 1958-1994sales, investment, capital Data are from Annual Surveystock, costs, etc. and Census of Manufacturing.

Number of establishments and Census of Manufacturing 1963, 67, 72, 77, 82, 87, 92.by size groups

Number of firms Census of Manufacturing 1963, 67, 72, 77, 82, 87, 92.

Four-firm concentration Census of Manufacturing 1963, 67, 72, 77, 82, 87, 92.

Used capital expenditures Census of Manufacturing 1972, 82, 92.

Rental payments Census of Manufacturing 1972, 82, 92.

Depreciation payments Census of Manufacturing 1972, 82, 92.

Aggregate variables: Economic Report of the 1958-1994.unemployment rate, federal President.funds rate, energy prices.

29

Table 1. Within-Year Across-Industries: Industry Structure Summary Statistics

Number of Establishments by Size Class

Large ö Smaller ö

Year Size: All Size: $500 Size: <500 Size: <250 Size: <100 Size: <50 FIRMS CONC

1963 639.9 10.8 629.1 614.3 572.0 518.1 587.9 37.9(1080) (17) (1079) (1072) (1037) (972) (1031) (21)

1967 612.9 12.5 600.4 583.4 539.0 480.6 555.4 38.2(976) (18) (975) (968) (933) (862) (922) (20)

1972 597.1 12.1 584.9 567.5 522.2 465.9 530.8 38.5(880) (17) (879) (872) (837) (773) (819) (20)

1977 644.5 12.1 632.4 614.5 568.2 512.1 574.6 38.3(917) (17) (916) (909) (876) (817) (869) (20)

1982 620.6 10.5 610.1 593.6 549.5 494.1 550.5 38.3(858) (15) (857) (852) (825) (772) (793) (20)

1987 614.1 9.3 604.8 589.1 545.6 490.8 545.8 39.8(875) (13) (874) (869) (840) (785) (807) (21)

1992 630.7 8.8 621.8 606.7 564.7 511.8 559.9 40.9(907) (13) (906) (901) (874) (823) (828) (21)

Notes:1. The data cover 267 SIC 4-digit U.S. manufacturing industries over the seven Census years 1963-1992.2. The numbers are the cross-industry mean value of the relevant industry structure variable; the corresponding standard deviations are in parentheses. Forexample, for 1992 the representative industry had about 600 firms and the s.d. of the number of firms was 828.

30

Table 2. Within-Industry Across-Years: Industry Structure Summary Statistics

Mean Std. Deviation Range: [Max.-Min.]

Size: All Mean 622.8 895.8 8,005

Std. Deviation 138.3 233.9 2,407

Size: $500 Mean 10.8 15.2 110

Std. Deviation 3.5 4.5 40

Size: <500 Mean 611.9 895.2 7,998

Std. Deviation 137.7 233.3 2,397

Size: <250 Mean 595.6 889.2 7,949

Std. Deviation 136.4 232.3 2,393

Size: <100 Mean 551.6 858.1 7,680

Std. Deviation 130.3 228.2 2,415

Size: <50 Mean 496.2 799.1 7,211

Std. Deviation 121.4 218.7 2,379

FIRMS Mean 557.8 834.7 7,556

Std. Deviation 129.2 234.3 2,548

CONC Mean 38.8 20.0 87

Std. Deviation 5.9 3.6 22Notes:1. The data cover 267 SIC 4-digit U.S. manufacturing industries over the seven Census years 1963-92.2. Row labeled “Mean”: For each industry we computed the “within-industry” mean value of the relevant industrystructure variable; we get 267 observations. This row presents the summary statistics for these means. For example,over the Census years 1963-92, the representative industry had about 558 firms and, for this variable, the standarddeviation was about 835.3. Row labeled “Std. Deviation”: For each industry we computed the “within-industry” standard deviation (s.d.)for the relevant industry structure variable. This column presents summary statistics for these s.d.’s. For example,for the number of firms the representative industry had a s.d. of about 129 and, for this variable, the standarddeviation was 234.

31

Table 3. Sunk Cost Summary Statistics

Median Mean Std. Dev.

USED 0.0795 0.0853 0.0454

RENT 0.0180 0.0269 0.0284

DEPR 0.0558 0.0577 0.0149

SUNK(EC) 0.0055 0.0137 0.0602Notes:1. USED is the share of used capital expenditures.2. RENT is the share of rental capital expenditures.3. DEPR is the share of depreciation expenditures.4. SUNK(EC) is sunk entry capital requirements.

32

Table 4. Within-Year Across-Industries: Explanatory Variables Summary Statistics

Period F(A) TECH(1) A GROW

1958-62 0.0175 0.0081 0.2425 0.0223(0.0140) (0.0226) (0.0881) (0.0122)

1963-66 0.0203 0.0088 0.2576 0.0263(0.0120) (0.0190) (0.0876) (0.0097)

1967-71 0.0213 0.0013 0.2681 0.0309(0.0127) (0.0212) (0.0851) (0.0118)

1972-76 0.0289 0.0073 0.2731 0.0377(0.0184) (0.0265) (0.0809) (0.0136)

1977-81 0.0239 0.0033 0.2757 0.0549(0.0143) (0.0257) (0.0838) (0.0214)

1982-86 0.0275 0.0046 0.2832 0.0584(0.0155) (0.0243) (0.0922) (0.0228)

1987-91 0.0262 0.0055 0.3086 0.0672(0.0190) (0.0241) (0.1005) (0.0270)

Notes:1. The data cover 267 SIC 4-digit U.S. manufacturing industries over the period 1958-94.2. The numbers are the cross-industry mean value of the relevant variable; the corresponding standard deviations are in parentheses. For example, for F(A)the representative industry had value of 0.0175 and the s.d. of F(A) was 0.014.

33

Table 5. Within-Industry Across-Years: Explanatory Variables Summary Statistics

Mean Std. Deviation Range: [Max.-Min.]

F(A) Mean 0.0236 0.0094 0.0603

Std. Deviation 0.0117 0.0072 0.0623

TECH(1) Mean 0.0070 0.0106 0.0893

Std. Deviation 0.0205 0.0099 0.0716

A Mean 0.2727 0.0825 0.5216

Std. Deviation 0.0358 0.0185 0.1142

GROW Mean 0.0425 0.0117 0.1177

Std. Deviation 0.0211 0.0091 0.0461Notes:1. The data cover 267 SIC 4-digit U.S. manufacturing industries over the period 1958-94.2. Row labeled “Mean”: For each industry we computed the “within-industry” mean value of the relevant variable;we get 267 observations. This row presents the summary statistics for these means. For example, over 1958-94, therepresentative industry had a F(A) value of 0.0236 and, for this variable, the standard deviation was about 0.0096.3. Row labeled “Std. Deviation”: For each industry we computed the “within-industry” standard deviation (s.d.)for the relevant variable. This column presents summary statistics for these s.d.’s. For example, for F(A) therepresentative industry had a s.d. of about 0.0117 and, for this variable, the standard deviation is 0.0072.

34

Table 6. Panel Summary Statistics

Mean Standard Deviation Range: [Max.-Min.]

Size: All 622.8 929.1 12,182

Size: $500 10.9 16.1 139

Size: <500 611.9 928.2 12,159

Size: <250 595.6 922.1 12,101

Size: <100 551.6 890.5 11,868

Size: <50 496.2 830.7 11,329

FIRMS 557.8 869.3 11,926

CONC 38.8 20.9 94

F(A) 0.0236 0.0158 0.1969

TECH(1) 0.0070 0.0236 0.2772

A 0.2727 0.0904 0.6321

GROW 0.0425 0.0242 0.1654Notes:1. Our dependent (industry structure) variables are the number of establishments and by size group - these aredenoted by “Size (.)”; number of firms, FIRMS; and industry four-firm output concentration, CONC. F(A) is profit-margin uncertainty, TECH(1) is the rate of technological change, A is industry profit-margin and GROW is theexpected growth of the industry as captured by the mean rate of new investment.

35

Table 7. Estimation Results for All Industries.Equation (7): STR ="+> F(A) +> TECH(1) +> A +> GROW +> ASTR +> STR +, .i,t i 1 i,t 2 i,t 3 i,t 4 i,t 5 t 6 i,t-1 i,t

Industry Structure Variable: STR

Number of Establishments

Size: All Size: $500 Size: <500 Size: <250 Size: <100 Size: <50 FIRMS CONC

F(A) -0.172** 0.093 -0.178** -0.209** -0.268** -0.308** -0.159** 0.191**i, t

(0.001) (0.258) (0.001) (0.001) (0.001) (0.001) (0.003) (0.005)

TECH(1) -1.737* 0.492 -1.809* -2.263** -2.943** -3.418** -1.642* 1.758i, t

(0.057) (0.729) (0.074) (0.046) (0.028) (0.015) (0.075) (0.193)

A 0.089 0.421* 0.029 -0.025 0.001 0.042 0.046 0.507**i, t

(0.504) (0.029) (0.849) (0.879) (0.995) (0.837) (0.725) (0.028)

GROW -0.041* -0.304** -0.017 -0.003 0.012 0.004 -0.011 0.058i, t

(0.094) (0.001) (0.521) (0.916) (0.726) (0.924) (0.678) (0.144)

ASTR 0.002** 0.001 0.002** 0.002** 0.003** 0.004** 0.001** 0.074**t

(0.001) (0.778) (0.001) (0.001) (0.002) (0.001) (0.037) (0.003)

STR 0.252** 0.261** 0.261** 0.253** 0.233** 0.208** 0.267** -0.049i, t-1

(0.001) (0.001) (0.001) (0.007) (0.005) (0.016) (0.001) (0.562)

Panel Obs. 1335 1335 1335 1335 1335 1335 1335 1335#Industries 267 267 267 267 267 267 267 267

Notes:1. As noted in section 5, the variables STR, F(A), A, GROW and ASTR in equation (7) are measured in logarithms; thus these coefficients can be interpretedas elasticities. TECH(1) is not measured in logarithms; thus the magnitude of these coefficients cannot be directly compared to others. For variabledefinitions, see notes to Table 6 and section 4.2. Specification (7) was estimated using the instrumental variables method; the instruments are described in section 5.3. p-values (two-tailed test) computed from heteroscedasticity-consistent standard errors are in parentheses; ** and * indicate significance at least at the 5%and 10% levels.4. For all the columns, the Hausman test easily rejected the null (at least at the 1% level) that the industry-specific variables were exogenous.

36

Table 8: Estimation Results by Sunk Cost Sub-Samples. Only the Uncertainty Coefficients are Reported.Equation (7): STR ="+> F(A) +> TECH(1) +> A +> GROW +> ASTR +> STR +, .i,t i 1 i,t 2 i,t 3 i,t 4 i,t 5 t 6 i,t-1 i,t

Sunk Cost Sub-Samples

ALL USED RENT DEPR SUNK(EC)Industries

Low Sunk High Sunk Low Sunk High Sunk Low Sunk High Sunk Low Sunk High Sunk

Size: All -0.172** -0.042 -0.175** 0.052 -0.306** 0.008 -0.289** -0.095 -0.172**(0.001) (0.580) (0.030) (0.467) (0.001) (0.910) (0.001) (0.134) (0.022)

Size: $500 0.093 0.052 0.041 0.077 0.100 0.197 -0.188* 0.173 -0.035(0.258) (0.647) (0.709) (0.526) (0.383) (0.183) (0.061) (0.130) (0.771)

Size: <500 -0.178** -0.052 -0.156* 0.055 -0.331** 0.005 -0.287** -0.099 -0.175**(0.002) (0.498) (0.092) (0.451) (0.001) (0.942) (0.001) (0.124) (0.035)

Size: <250 -0.209** -0.067 -0.198* 0.047 -0.391** -0.006 -0.327** -0.110* -0.230**(0.001) (0.395) (0.066) (0.523) (0.001) (0.936) (0.001) (0.094) (0.018)

Size: <100 -0.268** -0.087 -0.285** 0.033 -0.491** -0.017 -0.400** -0.137* -0.291**(0.001) (0.300) (0.025) (0.663) (0.001) (0.833) (0.001) (0.053) (0.013)

Size: <50 -0.308** -0.112 -0.312** 0.007 -0.553** -0.041 -0.464** -0.147* -0.353**(0.001) (0.213) (0.029) (0.927) (0.001) (0.664) (0.001) (0.060) (0.008)

FIRMS -0.159** -0.053 -0.163* 0.033 -0.298** 0.007 -0.281** -0.099 -0.144*(0.003) (0.476) (0.065) (0.635) (0.001) (0.926) (0.001) (0.125) (0.074)

CONC 0.191** 0.027 0.215** 0.097 0.278** 0.142 0.203** 0.009 0.214**(0.005) (0.778) (0.038) (0.257) (0.009) (0.206) (0.017) (0.903) (0.047)

Panel Obs. 1869 938 931 938 931 938 931 931 938#Industries 267 134 133 134 133 134 133 133 134

Notes:1. We estimated equation (7) for each sub-sample. See Table 7 for details. Only the uncertainty coefficients are presented.2. p-values (two-tailed test) computed from heteroscedasticity-consistent standard errors are in parentheses; ** and * indicate significance at least at the 5%and 10% levels.3. USED, RENT or DEPR greater than 50 percentile constitutes the low sunk cost samples; high if these are less than 50 percentile. SUNK(EC) less thanth th

50 percentile forms the low sunk costs sample; high if it is greater than 50 percentile. See section 4.3 for details.th th

37

Table 9: Additional Sunk Cost Sub-Samples. Only the Uncertainty Coefficients are Reported.Equation (7): STR ="+> F(A) +> TECH(1) +> A +> GROW +> ASTR +> STR +, .i,t i 1 i,t 2 i,t 3 i,t 4 i,t 5 t 6 i,t-1 i,t

Sunk Cost Sub-Samples

ALL Industries A. Combination of USED, RENT B. Combination of USED, RENT,and DEPR. DEPR and SUNK(EC).

Low Sunk High Sunk Low Sunk High Sunk

Size: All -0.172** 0.138 -0.314** -0.003 -0.325**(0.001) (0.194) (0.007) (0.976) (0.016)

Size: $500 0.093 0.075 -0.110 0.090 -0.074(0.258) (0.708) (0.421) (0.621) (0.652)

Size: <500 -0.178** 0.135 -0.286* -0.005 -0.300*(0.002) (0.206) (0.062) (0.959) (0.091)

Size: <250 -0.209** 0.134 -0.354* -0.012 -0.389*(0.001) (0.212) (0.073) (0.903) (0.087)

Size: <100 -0.268** 0.127 -0.531** -0.027 -0.576**(0.001) (0.254) (0.017) (0.798) (0.029)

Size: <50 -0.308** 0.102 -0.622** -0.047 -0.665**(0.001) (0.377) (0.012) (0.674) (0.024)

FIRMS -0.159** 0.075 -0.332** -0.037 -0.334**(0.003) (0.436) (0.017) (0.690) (0.038)

CONC 0.191** 0.018 0.408* 0.108 0.485**(0.005) (0.882) (0.055) (0.339) (0.047)

Panel Obs. 1869 310 305 250 245#Industries 267 62 61 50 49

Notes: We estimated (7) for each sub-sample. See Table 7 for details. Only the uncertainty coefficients are presented. In panel A, the combination “USEDand RENT and DEPR greater than 50 percentile” constitutes the low sunk cost sample; high if these are less than 50 percentile. In panel B, the combinationth th

“USED and RENT and DEPR greater than 50 percentile and SUNK(EC) less than 50 percentile” forms the low sunk cost sample; high otherwise.th th

38

Table 10. Additional Results for All Industries. Only the Uncertainty Coefficients are Reported.Equation (7): STR ="+> F(A) +> TECH(1) +> A +> GROW +> ASTR +> STR +, .i,t i 1 i,t 2 i,t 3 i,t 4 i,t 5 t 6 i,t-1 i,t

Number of Establishments

Size: All Size: $500 Size: <500 Size: <250 Size: <100 Size: <50 FIRMS CONC

Panel A: Estimates from Table 7.

-0.172** 0.093 -0.178** -0.209** -0.268** -0.308** -0.159** 0.191**(0.001) (0.258) (0.001) (0.001) (0.001) (0.001) (0.003) (0.005)

Panel B: Uncertainty constructed from profit forecasting equation:A =8 +8 A +8 A +8 SALES +8 SALES +8 FFR +8 FFR +8 EN +8 EN +,i,t 0 1 i,t-1 2 i,t-2 3 i,t-1 4 i,t-2 5 t-1 6 t-2 7 t-1 8 t-2 i,t

-0.169** 0.078 -0.178** -0.206** -0.258** -0.297** -0.152** 0.179**(0.002) (0.346) (0.002) (0.001) (0.001) (0.001) (0.006) (0.010)

Panel C: Uncertainty constructed from an AR(2) profit forecasting equation:A =8 +8 A +8 A +,i,t 0 1 i,t-1 2 i,t-2 i,t

-0.175** 0.106 -0.176** -0.208** -0.267** -0.316** -0.164** 0.146**(0.002) (0.188) (0.003) (0.002) (0.001) (0.001) (0.004) (0.043)

Panel D: Same as in panel A, but the profit-margin measure accounts for depreciation expenses:A(alt)=[(Total Sales Revenue-Total Variable Costs-Depreciation Expenses)/(Total Sales Revenue)].

-0.170** 0.071 -0.176** -0.203** -0.263** -0.294** -0.138** 0.158**(0.001) (0.341) (0.001) (0.001) (0.001) (0.001) (0.004) (0.010)

Panel E: Same as in Panel A, but GROW in equation (7) is growth of sales instead of the rate of new investment.

-0.189** 0.058 -0.182** -0.200** -0.254** -0.291** -0.232** 0.236**(0.001) (0.484) (0.002) (0.001) (0.001) (0.001) (0.001) (0.001)

(PH/MH)=(Entry price/Full cost)

(PL/ML)=(Exit price/Variable cost)

Sunk cost

Ratio

Figure 1(a). Increase in sunk cost and the entry/exit trigger

1.0

Uncertainty

Ratio

1.0

(PH/MH)=(Entry price/Full cost)

(PL/ML)=(Exit price/Variable cost)

Figure 1(b). Increase in uncertainty and the entry/exit trigger

39

40

< This page contains Figures 2(a)-2(g): The Establishment Size Distribution >< The graphs are included in the hard copy >